Detection of tele-local seismic phases by convolutional neural network and model interpretation
-
摘要: 利用北京国家观象台的测震记录，探索了样本构建、训练过程、模型结构等因素对远震震相P-S和近震震相Pg-Sg拾取模型性能的影响。结果表明：适中的卷积层深度、正则化和数据清洗能够有效地改善模型性能，而残差块的影响却相对有限。与此同时，基于类模型可视化和平滑GradCAM++的模型解释显示：卷积神经网络复现了震相的关键特征，其决策敏感区域也与震相识别的经验准则一致。最后，连续波形的扫描结果展示了卷积神经网络在远-近地震震相识别的应用前景与提升空间。此外，本文针对模型搭建与训练中存在的问题提出了样本选择、模型架构、标签标注和集成学习等改进方案，以供后续研究参考。Abstract: A better knowledge about the interrelationship between convolutional neural network （CNN） performance and its sample selection, training procedure, structure, etc., will be beneficial to employ this technique efficiently. We use CNN to detect tele-seismic P-S and local seismic Pg-Sg phases recorded by the Beijing National Earth Observatory. From the results by different parameter associations, it shows that the moderate layer depth, proper regularization, and data wash can significantly enhance the CNN performance while residual blocks giving only marginal improvement. Furthermore, we employ class model visualization and smooth GradCAM++ techniques to interpret the optimal CNN model. The results show that our model has learned the fundamental features of the seismic phases, with decision-sensitive distribution agreeing well with a priori knowledge. Also we use CNN model to scan the continuous seismic waveform, which exhibits its potentiality in seismic phase detection. Lastly, topics on sample selection, model framework, sample labelling, and ensemble learning are discussed for further work.
-
-
![]()
图 14 正则化对各深度模型的影响
（a） 正则化系数λ、模型层数与模型准确率之间的关系，圆点对应最高模型准确率；（b） 正则化系数λλ与模型均方之间的关系，其中空心图形对应最高准确率模型（注意2层数据点叠覆于3层之下）；（c） 模型损失函数的构成
Figure 14. The effects of regularization on CNN with different depths
（a） Relationship among regularization factor λ，CNN depth，and accuracy；（b） Relationship between regularization factor λ and the squaremean of model weights，with hollow patterns standing for CNN with highest accuracy （2-layer dot is overlapped by that of 3-layer）；（c） Variation of loss function with CNN depth
![]()
图 1 远震震中与震相数据的分布
（a）台站记录的远震位置（上）及 P 和 S 震相信噪比随震中距的分布（忽略了 35 个震中距异常的事件）（下）；（b）近震 Pg-Sg 走时差 40 s 内统计直方图（上）和 Pg-Sg 震相走时差与信噪比SNR分布（下）
Figure 1. Distribution of teleseismic epicenters and seismic data
（a） Locations of observatory and teleseismic epicenters (top）；Distribution of signal noise ratios of P and S phases with epicenter distances (bottom)；（b） Histogram of Pg-Sg travel time residuals within 40 s （top）；Distribution of signal noise ratios of Pg and Sg phases with Pg-Sg travel time residuals （bottom）
![]()
图 2 卷积神经网络（a）和瓶颈式残差块（b）结构示意图
红色虚线对应于层深为4的CNN结构
Figure 2. The structure of the convolutional neural network （a） and a bottle-neck residual block （b）
The red dotted line corresponds to the CNN structure with a depth of 4 layers
![]()
图 4 卷积层深度对模型性能的影响
（a） 未清洗数据，第4轮训练中不同卷积层深度（线上序号）模型的表现；（b） 未清洗数据，各轮次训练中不同卷积层数模型的最高准确率和最低损失函数；（c） 已清洗数据，第10轮训练中不同卷积层深度（线上序号）模型的表现；（d） 已清洗数据，各轮次训练中不同卷积层数模型的最高准确率及最低损失值
Figure 4. The influence of convolutional layer depth on model performance
（a） Data unwashed，the model perfomance for different depths （marked by numbers） during the 4th training round；（b） Data unwashed，the maximum accuray and minimum loss with different depths in the total 10 training rounds；（c） Data washed，the model perfomance for different depths （marked by numbers） during the 10th training round；（d） Data washed，the maximum accuray and minimum loss with different depths in the total 10 training rounds
![]()
图 5 类模型可视化算法的伪代码（Nguass＝20）
Figure 5. Pseudo code of Class Model Visualisation （Ngauss＝20）
![]()
图 7 模型解释结果
（a） 类模型可视化（CMV）；（b） 平滑GradCAM++. 绿色背景为原始波形
Figure 7. Results of model interpretation
（a） Class model visualization （CMV）； （b） Smooth GradCAM++. The green background stands for the original waveforms
![]()
图 8 连续观测震相片段的走时测量结果
（a） 各震相走时偏差及模型得分的分布；（b） 各阈值下诸震相的识别数目（蓝色方点）与走时偏差（红色圆点）
Figure 8. Traveltime mearurements for seismic phase clips cut from continuous observation
（a） Travetime residuals and CNN-predicted phase scores；（b） Detection number （blue square） and traveltime residuals （red dot） of the phases
![]()
图 10 正确的震相识别
水平虚线由低到高依次对应长、短步长的得分阈值，原始波形经归一化处理，下同
Figure 10. Seismic phases detected correctly
Horizontal dotted lines correspond to thresholds for long （lower） and short （upper） scanning steps，the waveform is normalized，the same below
![]()
图 11 错误震相的识别
（a） P震相波形清晰；（b） P震相波形不清晰；（c） S震相波形清晰；（d） S震相波形不清晰
Figure 11. Cases of the wrong detections
（a） Clear P waveform；（b） Unclear P waveform；（c） Clear S waveform；（d） Unclear S waveform
![]()
图 11 错误震相的识别
（e） Pg走时测量偏差过大；（f） Pg-Sg波形不清晰；（g） 疑似地震事件；（h） 搜索窗口分裂；（i） 后续震相（ScP）干扰；（j） 常见P震相误判波形；（k） 误判为S的ScS震相；（l） S震相的常见误判波形；（m） 未录入地震目录的Pg-Sg；（n） Pg-Sg常见误判波形
Figure 11. Cases of the wrong detections
（e） Too enormous Pg travetime residual；（f） Unclear Pg-Sg waveform；（g） Suspisious earthquake；（h） Splitting in searching window；（i） Later-coming phase （ScP）；（j） Common false P detection；（k） ScS identified as S；（l） Common false S detection； （m） Pg-Sg event not in catalogue；（n） Common false Pg-Sg detection
![]()
图 12 样本规模对模型准确率及样本利用情况的影响
（a）模型准确率与训练样本规模的关系；（b）训练样本规模与样本使用量的关系
Figure 12. Relationship among train data size，model accuracy，and samples used
（a） Relationship between train data size and model accuracy （b） Relationship between train data size and the number of samples used
![]()
图 13 各深度模型准确率与训练样本规模的关系
Figure 13. Relationship between the raining data size and model accuracy
![]()
图 15 不同模型的震相得分对比
（a） 模型选择；（b） 样本规模；（c） 正则化系数。黑色实线对应${y}＝{x}$，震相得分经单调变换${x}\to -\mathrm{l}\mathrm{g} （ 1－{x} ） $以便于展示
Figure 15. Phase scores for the selected models
（a） Model selection；（b） Training sample size；（c） L2 regularization factor. The black solid line corresponds to ${y}＝{x}$，with phase scores transformed monotonously by ${x}\to -\mathrm{l}\mathrm{g} （ 1－{x} ） $ for better view
-
李健,王晓明,张英海,王卫东,商杰,盖磊. 2020. 基于深度卷积神经网络的地震震相拾取方法研究[J]. 地球物理学报,63(4):1591–1606. doi: 10.6038/cjg2020N0057 Li J,Wang X M,Zhang Y H,Wang W D,Shang J,Ge L. 2020. Research on the seismic phase picking method based on the deep convolution neural network[J]. Chinese Journal of Geophysics,63(4):1591–1606 (in Chinese).
于子叶,储日升,盛敏汉,马海超. 2020. 兼顾速度和精度的深度神经网络震相拾取[J]. 地震学报,42(3):269–282. doi: 10.11939/jass.20190154 Yu Z Y,Chu R S,Sheng M H,Ma H C. 2020. A new deep neural network for phase picking with balanced speed and accuracy[J]. Acta Seismologica Sinica,42(3):269–282 (in Chinese).
赵明,陈石,Yuen D. 2019a. 基于深度学习卷积神经网络的地震波形自动分类与识别[J]. 地球物理学报,62(1):374–382. Zhao M,Chen S,Yuen D. 2019a. Waveform classification and seismic recognition by convolution neural network[J]. Chinese Journal of Geophysics,62(1):374–382 (in Chinese).
赵明,陈石,房立华,Yuen D A. 2019b. 基于U形卷积神经网络的震相识别与到时拾取方法研究[J]. 地球物理学报,62(8):3034–3042. Zhao M,Chen S,Fang L H,Yuen D A. 2019b. Earthquake phase arrival auto-picking based on U-shaped convolutional neural network[J]. Chinese Journal of Geophysics,62(8):3034–3042 (in Chinese).
中国地震台网中心. 2020. 历史查询[DB/OL]. [2020-05-02]. http://www.ceic.ac.cn/history. China Earthquake Networks Center. 2020. History querying[DB/OL]. [2020-05-02]. http://www.ceic.ac.cn/history （in Chinese）.
周本伟,范莉苹,张龙,李珀任,房立华. 2020. 利用卷积神经网络检测地震的方法与优化[J]. 地震学报,42(6):669–683. Zhou B W,Fan L P,Zhang L,Li P R,Fang L H. 2020. Earthquake detection using convolutional neural network and its optimization[J]. Acta Seismologica Sinica,42(6):669–683 (in Chinese).
Allen R. 1982. Automatic phase pickers:Their present use and future prospects[J]. Bull Seismol Soc Am,72(6B):S225–S242. doi: 10.1785/BSSA07206B0225
Chattopadhyay A, Sarkar A, Howlader P, Balasubramanian V N. 2018. Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks[C]//2018 IEEE Winter Conference on Applications of Computer Vision （WACV）. Lake Tahoe: IEEE: 839–847.
He K M, Zhang X Y, Ren S Q, Sun J. 2015. Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification[C]//2015 IEEE International Conference on Computer Vision （ICCV）. Santiago: IEEE: 1026–1034.
He K M, Zhang X Y, Ren S Q, Sun J. 2016. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. Las Vegas: IEEE: 770–778.
Kingma D P, Ba J. 2015. Adam: A method for stochastic optimization[C]//3rd International Conference on Learning Representations. San Diego: ICLR.
Mahendran A, Vedaldi A. 2015. Understanding deep image representations by inverting them[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. Boston: IEEE: 5188–5196.
Omeiza D, Speakman S, Cintas C, Weldermariam K. 2019. Smooth Grad-CAM++: An enhanced inference level visualization technique for deep convolutional neural network models[EB/OL]. [2019-08-03]. https://arxiv.org/pdf/1908.01224v1.pdf.
Ronneberger O, Fischer P, Brox T. 2015. U-net: Convolutional networks for biomedical image segmentation[C]//Proceeding of the 18th International Conference on Medical Image Computing and Computer-Assisted Intervention （MICCAI）. Munich: Springer: 234–241.
Ross Z E,White M C,Vernon F L,Ben-Zion Y. 2016. An improved algorithm for real-time S-wave picking with application to the (augmented) ANZA network in Southern California[J]. Bull Seismol Soc Am,106(5):2013–2022. doi: 10.1785/0120150230
Saragiotis C D,Hadjileontiadis L J,Rekanos I T,Panas S M. 2004. Automatic P phase picking using maximum kurtosis and κ-statistics criteria[J]. IEEE Geosci Remote Sens Lett,1(3):147–151. doi: 10.1109/LGRS.2004.828915
Selvaraju R R, Das A, Vedantam R, Cogswell M, Parikh D, Batra D. 2016. Grad-CAM: Why did you say that? Visual explanations from deep networks via gradient-based localization[EB/OL]. [2019-12-03]. https://arxiv.org/pdf/1610.02391v1.pdf.
Simonyan K, Vedaldi A, Zisserman A. 2014. Deep inside convolutional networks: Visualising image classification models and saliency maps[C]//2nd International Conference on Learning Representations. Banff: ICLR.
Sleeman R,Van Eck T. 1999. Robust automatic P-phase picking:An on-line implementation in the analysis of broadband seismogram recordings[J]. Phys Earth Planet In,113(1/2/3/4):265–275.
Springenberg T, Dosovitskiy A, Brox T, Riedmiller A. 2014. Striving for simplicity: The all convolutional net[EB/OL]. [2015-04-13]. https://arxiv.org/pdf/1412.6806.pdf.
USGS. 2020. Magnitude 4.5+ earthquakes, past Day[DB/OL]. [2020-05-02]. https://earthquake.usgs.gov/earthquakes/map/?extent=18.06231,-137.19727&extent=54.31652,-52.82227.
Zhou B L, Khosla A, Lapedriza G A, Oliva A, Torralba A. 2016. Learning deep features for discriminative localization[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. Las Vegas: IEEE: 2921–2929.
Zhu L J,Peng Z G,McClellan J,Li C Y,Yao D D,Li Z F,Fang L H. 2019. Deep learning for seismic phase detection and picking in the aftershock zone of 2008 MW7.9 Wenchuan earthquake[J]. Phys Earth Planet In,293:106261. doi: 10.1016/j.pepi.2019.05.004
Zhu W Q,Beroza G C. 2019. PhaseNet:A deep-neural-network-based seismic arrival-time picking method[J]. Geophys J Int,216(1):261–273.
-
其他相关附件
-
PDF格式
附件1 点击下载(754KB) -
PDF格式
附件2 点击下载(683KB) -
PDF格式
附件3 点击下载(2590KB)
-
下载:
计量
- 文章访问数: 436
- HTML全文浏览量: 175
- PDF下载量: 92
