Supervised Machine Learning of High Rate GNSS Velocities for Earthquake Strong Motion Signals
Supervised Machine Learning of High Rate GNSS Velocities for Earthquake Strong Motion Signals
复制标题
DOI：
10.1029/2022jb024854
复制
发表时间：
2022-10
期刊：
影响因子：
--
通讯作者：
T. Dittmann;Y. Liu;Y. Morton;D. Mencin
中科院分区：
文献类型：
--
作者：
T. Dittmann;Y. Liu;Y. Morton;D. Mencin
高速率全球导航卫星系统（GNSS）处理的时间序列捕捉到广泛的地震强震信号，但经历定期零星的噪音，可能难以区分真正的地震信号。在高的、位置变化的本底噪声中，可能的地震信号频率范围使得滤波难以推广。现有的自动检测方法依赖于外部输入来减轻错误警报，这限制了它们的有用性。出于这些原因，大地测量地震信号检测是数据驱动机器学习分类的一个令人信服的候选者。在这项研究中，我们产生了高速率GNSS时间差分载波相位（TDCP）速度时间序列并发的空间和时间的预期信号发生在近20年的77个地震。相对于传统的大地测量位移处理，TDCP速度处理提高了灵敏度，而不需要复杂的校正。我们训练、验证和测试了一个随机森林分类器，以区分地震事件和噪声。我们发现我们的监督随机森林分类器通过将频域和时域特征结合到决策标准中，在独立模式下优于现有的检测方法。该分类器在MW4.8-8.2的地震事件数据集内实现了90%的地震事件检测真阳性率，典型的检测延迟为S波到达后的几秒钟。我们的结论是，该模型的性能提供了足够的信心，使这些有价值的地面运动测量能够在独立模式下运行，用于开发边缘处理，大地测量基础设施监测和纳入业务地面运动观测和模型。
High rate Global Navigation Satellite System (GNSS) processed time series capture a broad spectrum of earthquake strong motion signals, but experience regular sporadic noise that can be difficult to distinguish from true seismic signals. The range of possible seismic signal frequencies amidst a high, location‐varying noise floor makes filtering difficult to generalize. Existing methods for automatic detection rely on external inputs to mitigate false alerts, which limit their usefulness. For these reasons, geodetic seismic signal detection makes for a compelling candidate for data‐driven machine learning classification. In this study we generated high rate GNSS time differenced carrier phase (TDCP) velocity time series concurrent in space and time with expected signals from 77 earthquakes occurring over nearly 20 years. TDCP velocity processing has increased sensitivity relative to traditional geodetic displacement processing without requiring sophisticated corrections. We trained, validated and tested a random forest classifier to differentiate seismic events from noise. We find our supervised random forest classifier outperforms the existing detection methods in stand‐alone mode by combining frequency and time domain features into decision criteria. The classifier achieves a 90% true positive rate of seismic event detection within the data set of events ranging from MW4.8–8.2, with typical detection latencies seconds behind S‐wave arrivals. We conclude the performance of this model provides sufficient confidence to enable these valuable ground motion measurements to run in stand‐alone mode for development of edge processing, geodetic infrastructure monitoring and inclusion in operational ground motion observations and models.