A Deep Convolutional Neural Network for Localization of Clustered Earthquakes Based on Multistation Full Waveforms
A Deep Convolutional Neural Network for Localization of Clustered Earthquakes Based on Multistation Full Waveforms
复制标题
DOI：
10.1785/0220180320
复制
发表时间：
2019-03-01
影响因子：
3.3
通讯作者：
Ohrnberger, Matthias
中科院分区：
文献类型：
--
作者：
Kriegerowski, Marius;Petersen, Gesa M.;Ohrnberger, Matthias
地震定位是地震学领域的一个必要条件，也是震源研究和灾害评估等进一步分析的先决条件。传统的定位方法通常依赖于手动拾取相位。我们提出了一种使用深度学习的替代方法，一旦经过训练，就可以有效地预测震源位置。在地震学中，神经网络通常使用单站记录或基于先前从波形中提取的特征进行训练。我们直接使用多个台站的三分量全波形记录。这意味着在预处理过程中不会丢失任何信息，并且数据的准备不需要专业知识。我们的深度卷积神经网络（CNN）的第一个卷积层对它所训练的波形的特征变得敏感。我们表明，这一层因此可以额外用作事件检测器。作为测试案例，我们使用来自西波西米亚的2000多个地震群事件训练CNN，这些事件由9个当地的三分量台站记录。CNN成功定位了908个验证事件，与双差重定位参考目录相比，东西方向的标准偏差为56.4米，南北方向的标准偏差为123.8米，垂直方向的标准偏差为136.3米。该检测器对幅度低至M-L = -0.8的事件敏感，假阳性检测率为3.5%。
Earthquake localization is both a necessity within the field of seismology, and a prerequisite for further analysis such as source studies and hazard assessment. Traditional localization methods often rely on manually picked phases. We present an alternative approach using deep learning that once trained can predict hypocenter locations efficiently. In seismology, neural networks have typically been trained with either single-station records or based on features that have been extracted previously from the waveforms. We use three-component full-waveform records of multiple stations directly. This means no information is lost during preprocessing and preparation of the data does not require expert knowledge. The first convolutional layer of our deep convolutional neural network (CNN) becomes sensitive to features that characterize the waveforms it is trained on. We show that this layer can therefore additionally be used as an event detector.As a test case, we trained our CNN using more than 2000 earthquake swarm events from West Bohemia, recorded by nine local three-component stations. The CNN successfully located 908 validation events with standard deviations of 56.4 m in east-west, 123.8 m in north-south, and 136.3 m in vertical direction compared to a double-difference relocated reference catalog. The detector is sensitive to events with magnitudes down to M-L = -0.8 with 3.5% false positive detections.