Development of a Heterogeneous Waveform Dataset and Evaluation of Phase Picking Models for Yunnan Earthquake Early Warning Network
-
摘要: 近年来，深度学习方法在地震检测和震相拾取中得到广泛的应用.然而现有模型主要基于高信噪比的速度型波形数据进行训练，缺乏对加速度计与烈度计数据的泛化性评估.为探究现有模型对加速度数据的处理效果及在云南地区的泛化能力，基于云南预警台网的最新观测数据，构建了包括速度计、加速度计和烈度计的多源异构高质量波形数据集，且所有震相到时均由人工标注.结合PhaseNet、USTC-Pickers等5种专业模型，以及SeisMoLLM和SeisT等4种大模型，系统评估了不同模型在云南数据集上的震相拾取性能.结果表明，本地迁移优化的USTC-Pickers综合性能最优，其Pg和Sg震相拾取的平均F1值达0.779（到时拾取差异△t≤0.1 s），显著优于其他模型，且在检测加速度计与烈度计数据时，较好解决了震相拾取滞后问题；大模型在Sg拾取等复杂环境中展现出更强的泛化能力.研究还揭示了主流地震检测模型在不同波形长度、震级、震中距条件下的性能变化，强调了本地化训练与模型选取在实际应用中的重要性.研究结果为地震预警系统中的地震检测和震相识别，以及中国地震科学实验场地震观测数据的实时自动处理提供参考.Abstract: In recent years, deep learning methods have been widely applied to seismic detection and phase picking. However, existing models are mainly trained on high signal-to-noise ratio (SNR) velocity-type waveform data, with limited evaluation of their generalization to accelerometer and intensity meter data. To investigate the performance of existing models on accelerometer data and their generalization capability in Yunnan, this study constructed a high-quality, multi-source heterogeneous waveform dataset based on the latest observations from the Yunnan Earthquake Early Warning (EEW) network, including velocity meters, accelerometers, and intensity meters, with all phase arrival times manually annotated. It systematically evaluated the phase-picking performance of nine models-on the Yunnan dataset five domain-specific models (e.g., PhaseNet, USTC-Pickers) and four large models (e.g., SeisMoLLM, SeisT). The locally fine-tuned USTC-Pickers achieved the best overall performance, with mean F1 scores of 0.779 for Pg and Sg phase picking (Δt≤0.1 s), significantly outperforming other models, and effectively mitigating phase-picking delays for accelerometer and intensity meter data. Large models demonstrated stronger generalization in Sg picking and low-SNR conditions. The study also revealed performance variations of mainstream seismic detection models under different waveform lengths, magnitudes, and epicentral distances, underscoring the importance of localized training and model architecture selection in practical applications. The research findings provide references for seismic detection and phase picking in earthquake early warning systems, as well as for the real-time automatic processing of seismic data at the China Earthquake Science Experiment Site.
-
Key words:
- seismic phase picking /
- dataset /
- large language model /
- performance evaluation /
- earthquake early warning /
- seismology
-
图 1 不同类型仪器记录的地震波形和PhaseNet检测的到时对比
云南预警台网SYX02台（同址安装速度计和加速度计）和距离其最近的J2811台（安装烈度计）记录到的2023年9月3日云南耿马ML3.4地震波形.黑色实线为归一化后的地震波形，蓝色实线为人工标记的Pg到时、红色实线为人工标记的Sg到时，蓝色虚线为PhaseNet拾取的Pg到时，红色虚线为PhaseNet拾取的Sg到时；图a为速度计，图b为加速度计，图c为烈度计.从图中可以看出，PhaseNet在速度计数据上检测的震相到时与人工标注的震相到时较为一致，在加速度计或者烈度计数据上检测的震相到时与人工标注的震相到时存在一定差异
Fig. 1. Comparison of seismic waveforms recorded by different instruments and PhaseNet-detected arrival times
图 2 预警工程建设前后云南地震台网密度及台站分布
a. 预警工程前速度计分布；b. 预警工程后速度计分布；c. 预警工程后速度计+加速度计分布；d.预警工程后速度计+加速度计+烈度计分布
Fig. 2. Station density and distribution of the Yunnan Seismic Network before and after the earthquake early warning project construction
图 3 数据集中地震事件震中分布和台站分布
红色圆圈代表不同震级地震事件，黑色三角形代表速度计，蓝色三角形代表加速度计，黄色三角形代表烈度计
Fig. 3. Epicenter and station distribution of earthquakes in the dataset
图 4 数据集特征统计图
a.震级统计；b.震中距；c.信噪比；d.速度计数据震相走时曲线；e.加速度计数据震相走时曲线；f.烈度计数据震相走时曲线. 图d、e、f中蓝色为Pg，红色为Sg，散点颜色深浅代表不同震级，颜色越深震级越大
Fig. 4. Statistical characteristics of the dataset
图 5 不同模型在三类数据上的精确率、召回率和F1雷达图
a.速度计；b.加速度计；c.烈度计
Fig. 5. Precision, recall, and F1-score radar chart of different models on three data types
图 6 不同模型在三类数据上的震相到时误差标准差及平均绝对误差对比
a.速度计；b.加速度计；c.烈度计
Fig. 6. Comparison of standard deviation and mean absolute error of phase arrival times for different models on three data types
图 7 不同模型在三类数据上的Pg震相到时误差(Tpre-Ttrue)分布直方图
Fig. 7. Histograms of Pg arrival time errors (Tpre-Ttrue) for different models on three data types
图 8 不同模型在三类数据上的Sg震相到时误差(Tpre-Ttrue)分布直方图
Fig. 8. Histograms of Sg arrival time errors (Tpre-Ttrue) for different models on three data types
图 9 模型在同一波形不同时窗长度数据上的拾取效果对比
左图为30 s窗长，右图为12 s窗长；图a、f为三分量原始波形，图b、g为真实标签到时扩展为高斯窗的概率曲线，图c和图h、图d和图i、图e和图j分别为PhaseNet、USTC-Pickers和SeisT输出的概率曲线
Fig. 9. Comparison of picking performance of different models on the same waveform with different time window lengths
图 10 模型在不同信噪比、震级和震中距离的速度波形上的拾取效果对比
Fig. 10. Phase picking performance on velocity data under different SNR, magnitude, and epicentral distance
图 11 模型在不同信噪比、震级和震中距的加速度计数据的拾取效果对比
Fig. 11. Phase picking performance on accelerogram under different SNR, magnitude, and epicentral distance
图 12 模型在不同信噪比、震级和震中距的烈度计数据的拾取效果对比
Fig. 12. Phase picking performance on MEMS data under different SNR, magnitude, and epicentral distance
图 13 PhaseNet对同台址速度计和加速度计震相拾取对比
a. P和S的震相数量；b. 速度计与加速度计P到时误差T（HH）-T（HN）分布；c. 速度计与加速度计S到时误差T（HH）-T（HN）分布
Fig. 13. Phase picking comparison on co-located velocity and acceleration data
图 14 各模型在异构数据集上的F1均值与推理速度对比
Fig. 14. Comparison of average F1-score and inference speed of different models on heterogeneous datasets
表 1 云南地区部署的地震计类型
Table 1. Summary of seismometer types deployed in Yunnan region
台站 背景噪声(m/s2) 类型 设备型号 频带 灵敏度 通道 数量 制造商 基准站 1.0×10‒9~1.0×10‒8 速度计 GL_CS60 60 s~50 Hz 2 000a HH* 78 港震 GL_CS120 120 s~50 Hz 2 000a HH* 33 港震 ITC-60A 60 s~50 Hz 2 000a HH* 43 天元 ITC-120A 120 s~50 Hz 2 000a HH* 31 天元 BBVS-60 60 s~50 Hz 2 000a HH* 10 港震 JS-60 60 s~50 Hz 2 000a HH* 4 深研院 JS-120 120 s~50 Hz 2 000a HH* 3 深研院 加速度计 JS-A2 DC~80 Hz 2.5b HN* 101 深研院 TDA-33M DC~80 Hz 2.5b HN* 101 泰德 基本站 1.0×10‒6~1.0×10‒5 加速度计 TDA-33M DC~80 Hz 2.5b HN* 113 泰德 JS-A2 DC~80 Hz 2.5b HN* 115 深研院 一般站 1.0×10‒4~1.0×10‒3 烈度计 GL-P2B DC~80 Hz 1.0×10c6 EI* 20 港震 Palert Advance DC~80 Hz - EI* 100 勤联 TMA-33 DC~80 Hz - EI* 705 泰德 VH-GL-LDY DC~80 Hz 1.0×10c6 EI* 405 瑞琪 注：a.单位为V·m-1·s; b.单位为V·m-1·s2; c.单位为count·m-1·s2. 
下载: 导出CSV
表 2 地震检测模型信息
Table 2. Summary of information on earthquake detection models
模型类型 模型名称 训练数据集 训练策略 参考文献 专业模型 PhaseNet NCEDC 全监督训练 Zhu and Beroza, 2019 EQTransformer STEAD 全监督训练 Mousavi et al., 2020 RNN CSNCD 全监督训练 Yu et al., 2023 LPPN STEAD 全监督训练 Yu and Wang, 2022 USTC-Pickers DiTing 预训练+增量微调 Zhu et al., 2023 大模型 SeisT STEAD/DiTing 联合训练, 多任务优化 Li et al., 2024 PRIME_DP CSNCD 预训练+解码器微调 Yu et al., 2024 SeisMoLLM STEAD/DiTing 跨模态微调 Wang et al., 2025 SeisLM STEAD 自监督预训练+微调 Liu et al., 2024
下载: 导出CSV
表 3 不同模型在本文数据集上的拾取结果统计(误差在0.1 s为TP)
Table 3. Summary of picking results of different models on the dataset (error within 0.1 s counted as TP)
测试数据 测试模型 Pg Sg Precision Recall F1 Precision Recall F1 速度计 PhaseNet 0.827 0.867 0.847 0.528 0.577 0.551 EQTransformer 0.859 0.870 0.864 0.702 0.631 0.665 RNN 0.842 0.822 0.832 0.434 0.454 0.444 LPPN 0.801 0.765 0.782 0.576 0.413 0.481 USTC-Pickers 0.870 0.879 0.874 0.637 0.688 0.662 SeisT 0.858 0.856 0.857 0.663 0.634 0.648 PRIME_DP 0.806 0.803 0.805 0.479 0.539 0.507 SeisMoLLM 0.804 0.865 0.833 0.653 0.681 0.666 SeisLM 0.863 0.889 0.876 0.579 0.658 0.616 平均值 0.837 0.846 0.841 0.584 0.586 0.582 加速度计 PhaseNet 0.802 0.846 0.823 0.422 0.565 0.484 EQTransformer 0.876 0.832 0.853 0.579 0.622 0.600 RNN 0.720 0.691 0.705 0.399 0.494 0.441 LPPN 0.676 0.595 0.633 0.501 0.440 0.468 USTC-Pickers 0.936 0.935 0.936 0.588 0.753 0.660 SeisT 0.918 0.904 0.911 0.605 0.701 0.650 PRIME_DP 0.760 0.735 0.747 0.393 0.525 0.449 SeisMoLLM 0.880 0.858 0.869 0.568 0.711 0.632 SeisLM 0.859 0.903 0.881 0.475 0.667 0.555 平均值 0.825 0.811 0.818 0.503 0.609 0.549 烈度计 PhaseNet 0.810 0.831 0.820 0.479 0.611 0.537 EQTransformer 0.900 0.617 0.732 0.665 0.529 0.589 RNN 0.779 0.724 0.750 0.499 0.590 0.541 LPPN 0.663 0.474 0.553 0.554 0.464 0.505 USTC-Pickers 0.827 0.819 0.823 0.648 0.803 0.718 SeisT 0.883 0.828 0.855 0.682 0.785 0.730 PRIME_DP 0.766 0.602 0.674 0.471 0.533 0.500 SeisMoLLM 0.951 0.535 0.684 0.617 0.698 0.655 SeisLM 0.865 0.776 0.818 0.541 0.701 0.611 平均值 0.827 0.689 0.746 0.573 0.635 0.598
下载: 导出CSV
-
An, Y. R., 2024. Introduction to a Recently Released Dataset Entitled CSNCD: A Comprehensive Dataset of Chinese Seismic Network. Earthquake Research Advances, 4(1): 100255. https://doi.org/10.1016/j.eqrea.2023.100255 Bormann, P., Wendt, S., Klinge, K., 2013. Data Analysis and Seismogram Interpretation. New Manual of Seismological Observatory Practice 2 (NMSOP-2), 11: 1-34. https://doi.org/10.2312/GFZ.NMSOP-2_CH11 Bornstein, T., Lange, D., Münchmeyer, J., et al., 2024. PickBlue: Seismic Phase Picking for Ocean Bottom Seismometers with Deep Learning. Earth and Space Science, 11(1): e2023EA003332. https://doi.org/10.1029/2023EA003332 Chai, C. P., Maceira, M., Santos-Villalobos, H. J., et al., 2020. Using a Deep Neural Network and Transfer Learning to Bridge Scales for Seismic Phase Picking. Geophysical Research Letters, 47(16): e2020GL088651. https://doi.org/10.1029/2020GL088651 Du, M. N., He, F. X., Zou, N., et al., 2024. Shortcut Learning of Large Language Models in Natural Language Understanding. Communications of the ACM, 67(1): 110-120. https://doi.org/10.1145/3596490 Fang, L. H., Wu, Z. L., Song, K., 2017. SeismOlympics. Seismological Research Letters, 88(6): 1429-1430. https://doi.org/10.1785/0220170134 Feng, C. Y., Yang, Y., Hu, X., et al., 2023. MSNet: A Seismic Phase Picking Network Applicable to Microseismic Monitoring. IEEE Geoscience and Remote Sensing Letters, 20: 1-5. https://doi.org/10.1109/LGRS.2023.3314029 Huangfu, G., 2009. Research on the Seismicity in Yunnan, China (Dissertation). University of Science and Technology of China, Hefei (in Chinese with English abstract). Li, J., Chen, J. Y., Tang, Y. C., et al., 2023. Transforming Medical Imaging with Transformers? A Comparative Review of Key Properties, Current Progresses, and Future Perspectives. Medical Image Analysis, 85: 102762. https://doi.org/10.1016/j.media.2023.102762 Li, S., Yang, X., Cao, A. Y., et al., 2024. SeisT: A Foundational Deep-Learning Model for Earthquake Monitoring Tasks. IEEE Transactions on Geoscience and Remote Sensing, 62: 1-15. https://doi.org/10.1109/TGRS.2024.3371503 Liu, T. L., Münchmeyer, J., Laurenti, L., et al., 2024. SeisLM: A Foundation Model for Seismic Waveforms. arXiv: 2410.15765. https://doi.org/10.48550/arXiv.2410.15765 Mousavi, S. M., Beroza, G. C., 2022. Deep-Learning Seismology. Science, 377(6607): eabm4470. https://doi.org/10.1126/science.abm4470 Mousavi, S. M., Ellsworth, W. L., Zhu, W. Q., et al., 2020. Earthquake Transformer—An Attentive Deep-Learning Model for Simultaneous Earthquake Detection and Phase Picking. Nature Communications, 11: 3952. https://doi.org/10.1038/s41467-020-17591-w Mousavi, S. M., Sheng, Y. X., Zhu, W. Q., et al., 2019. STanford EArthquake Dataset (STEAD): A Global Data Set of Seismic Signals for AI. IEEE Access, 7: 179464-179476. https://doi.org/10.1109/ACCESS.2019.2947848 Münchmeyer, J., Woollam, J., Rietbrock, A., et al., 2022. Which Picker Fits My Data? A Quantitative Evaluation of Deep Learning Based Seismic Pickers. Journal of Geophysical Research: Solid Earth, 127(1): e2021JB023499. https://doi.org/10.1029/2021JB023499 Niksejel, A., Zhang, M., 2024. OBSTransformer: A Deep-Learning Seismic Phase Picker for OBS Data Using Automated Labelling and Transfer Learning. Geophysical Journal International, 237(1): 485-505. https://doi.org/10.1093/gji/ggae049 Pecher, B., Srba, I., Bielikova, M., 2025. Comparing Specialised Small and General Large Language Models on Text Classification: 100 Labelled Samples to Achieve Break-Even Performance. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, Suzhou, 165-184. https://10.18653/v1/2025.emnlp-main.9 Peng, C. Y., Jiang, P., Ma, Q., et al., 2022. Chinese Nationwide Earthquake Early Warning System and Its Performance in the 2022 Lushan M6.1 Earthquake. Remote Sensing, 14(17): 4269. https://doi.org/10.3390/rs14174269 Ross, Z. E., Meier, M. A., Hauksson, E., et al., 2018. Generalized Seismic Phase Detection with Deep Learning. Bulletin of the Seismological Society of America, 108(5A): 2894-2901. https://doi.org/10.1785/0120180080 Saad, O. M., Savvaidis, A., Chen, Y. K., 2024. Transfer Learning for Seismic Phase Picking with Significantly Higher Precision in Faraway Seismic Stations. IEEE Transactions on Geoscience and Remote Sensing, 62: 1-12. https://doi.org/10.1109/TGRS.2024.3422612 Si, X., Wu, X. M., Li, Z. F., et al., 2024. An All-in-One Seismic Phase Picking, Location, and Association Network for Multi-Task Multi-Station Earthquake Monitoring. Communications Earth & Environment, 5: 1-13. https://doi.org/10.1038/s43247-023-01188-4 Wang, X. H., Liu, F., Su, R., et al., 2025. SeisMoLLM: Advancing Seismic Monitoring via Cross-Modal Transfer with Pre-Trained Large Language Model. arXiv: 2502.19960. https://doi.org/10.48550/arXiv.2502.19960 Woollam, J., Münchmeyer, J., Tilmann, F., et al., 2022. SeisBench-A Toolbox for Machine Learning in Seismology. Seismological Research Letters, 93(3): 1695-1709. https://doi.org/10.1785/0220210324 Yu, Z. Y., Cai, Y. Q., Wang, W. T., et al., 2024. PRIME-DP: Pre-Trained Integrated Model for Earthquake Data Processing. arXiv: 2408.01919. https://doi.org/10.48550/arXiv.2408.01919 Yu, Z. Y., Wang, W. T., 2022. LPPN: A Lightweight Network for Fast Phase Picking. Seismological Research Letters, 93(5): 2834-2846. https://doi.org/10.1785/0220210309 Yu, Z. Y., Wang, W. T., Chen, Y. N., 2023. Benchmark on the Accuracy and Efficiency of Several Neural Network Based Phase Pickers Using Datasets from China Seismic Network. Earthquake Science, 36(2): 113-131. https://doi.org/10.1016/j.eqs.2022.10.001 Zhao, M., Xiao, Z. W., Chen, S., et al., 2023. DiTing: A Large-Scale Chinese Seismic Benchmark Dataset for Artificial Intelligence in Seismology. Earthquake Science, 36(2): 84-94. https://doi.org/10.1016/j.eqs.2022.01.022 Zhu, J., Li, Z. F., Fang, L. H., 2023. USTC-Pickers: A Unified Set of Seismic Phase Pickers Transfer Learned for China. Earthquake Science, 36(2): 95-112. https://doi.org/10.1016/j.eqs.2023.03.001 Zhu, J., Zhong, Y. S., 2025. Evaluation of the Monitoring Capability of the Yunnan Earthquake Early Warning Network. Journal of Seismological Research, 48(3): 422-431 (in Chinese with English abstract). Zhu, W. Q., Beroza, G. C., 2019. PhaseNet: A Deep-Neural-Network-Based Seismic Arrival Time Picking Method. Geophysical Journal International, 216(1): 261-273. https://doi.org/10.1093/gji/ggy423 皇甫岗, 2009. 云南地震活动性研究(博士学位论文). 合肥: 中国科学技术大学. 朱杰, 钟玉盛, 2025. 云南地震预警站网监测效能评估. 地震研究, 48(3): 422-431. -
点击查看大图
计量
- 文章访问数: 134
- HTML全文浏览量: 5
- PDF下载量: 12
- 被引次数: 0
