|
引言：随着2017-2018年数据科学的升温，机器学习（Machine Learning）在地震学中的应用也逐渐增多。地震科学拥有海量的数据，机器学习在未来必然会在地震学中有一席之地。Seismological Research Letter开始推送Focus section on machine learning in seismology的版块，我们将为大家提供2019年一月的Early Publication中的五篇机器学习论文的导读，给大家带来最新的资讯。 1. 标题：Discrimination of Seismic Signals from Earthquakes and Tectonic Tremor by Applying a Convolutional Neural Network to Running Spectral Images. 作者：Masaru Nakano, D. Sugiyama, T. Hori, T. Kuwatani, S. Tsuboi. 期刊：Seismological Research Letter (SRL) 日期：2019年1月3日 链接：https:///10.1785/0220180279 摘要导读： 监测俯冲版块和走滑版块边界缓慢变形释放的地震信号（常称为构造震颤）有可能揭示逆冲带地震震源区的变形过程。震颤信号主要发生在2-8Hz的频段，并有可能持续几十秒至几分钟。这与典型地震产生的持续小于一分钟且频率高达几十赫兹的信号有显著不同。因为地震震颤源于随机过程，其波形由随机函数表示。这导致利用确定性方法区别震颤信号和地震信号十分困难。在这个研究中，我们利用卷积神经网络分析运行中的构造震颤和地震信号频谱，从而将震颤信号从本地地震信号中区分出来。我们发展了Seismic running spectra-CNN(SRSpec-CNN)的方法，它对反映了信号源物理属性的信号的绝对频率十分敏感，但对于信号起始时间不敏感。SRSpec-CNN是从17213个64*64像素的图片集中训练出来并有130211个参数。根据对信号频率和时长的联合分析，我们对构造震颤，地震信号和噪音识别达到了99.5%的准确率。由于运行中频谱能够明确区分这些信号的特征，我们能够利用简单的CNN达到如此高的准确率。 Abstract：Monitoring of seismic signals generated by slow deformation at convergent and transform plate boundaries worldwide, known as tectonic tremor, might provide insights into deformation processes in the source regions of megathrust earthquakes. Tremor signals occur dominantly in the 2–8 Hz frequency band and can last for tens of seconds to several minutes, in contrast to typical earthquakes that produce seismic signals at frequencies up to several tens of hertz and last less than a minute. Because tremor is caused by stochastic processes, the resultant waveforms are represented by a stochastic function and construction of deterministic measures to discriminate tremor signals from earthquakes is very difficult. In this study, we used a convolutional neural network (CNN) to discriminate the signals of tectonic tremor from those of local earthquakes in running spectral images of these signals. We developed a method (seismic running spectra‐CNN [SRSpec‐CNN]) that is sensitive to the absolute frequency of signal appearance, which reflects the physical properties of the signal source, but is insensitive to the time of signal onset. SRSpec‐CNN has 130,211 parameters that were trained by 17,213 images of 64×64 pixels. Based on simultaneous analyses of the frequency contents and durations of the signals, we achieved 99.5% accuracy for our identifications of signals from tectonic tremor, local earthquakes, and noise. Because running spectra clearly differentiate the characteristic features of these signals, we were able to achieve this high accuracy by using a CNN of simple architecture. 2. 标题：Application of Pool-based active learning in physics-based earthquake ground-motion simulation. 作者：Naeem Khoshnevis, Ricardo Taborda. 日期：2019年1月3日 链接：https:///10.1785/0220180296 摘要导读： 我们展示了应用主动学习（Active Learning）来发展代替基于物理模型的地震震动模拟器（physics-based earthquake ground-motion simulators)。这些代理（将输入参数和输出参数相关联但不需要高强度计算）将成为任何参数优化，敏感度和不确定分析的必要组成。人工智能神经网络（Artificial Nerual Networks, ANNs），作为这类代理的一个例子，能十分高效地估计复杂模型。ANNs利用一个训练数据集学习输入输出相互关联的过程。训练集是一套输入参数和对应的输出参数。产生训练集需要做很多区域地震动模拟。这些数值模拟计算上十分有挑战。所以，我们需要用一个步进（Step-by-Step）的学习方法来减少不必要训练集的模拟。我们称这些方法为主动学习。本研究中，我们利用pool-based query by committe (QBC) 主动学习方法，结合有效的初始化，来研究模型在训练过程中的表现。我们利用一个数据集。这个数据集是模拟一个区域地震的非弹性衰减，并主要估计地面峰值速度（Peak Ground Velocity）。结果表明当同样数量的训练集被用来训练主动学习和被动学习（Passive Learning），主动学习在减小一般化误差（generalization error）上表现更好。当训练集数据增加时，主动学习表现的变化显著地比较小，这说明主动学习方法的稳定和可预测性。这个研究，虽然只限于一个地震和一个判别标准，表明发展‘’基于物理模型的地震震动模拟器’的代理，主动学习的应用将是降低计算量和产生稳定的预测器的重要的一步。 Abstract: We present the results of the application of the active learning method in developing surrogates as physics‐based earthquake ground‐motion simulators. The surrogates, which map input parameters into output values without demanding intensive computations, are an essential part of any parameter optimization, sensitivity, and uncertainty analysis. Artificial neural networks (ANNs), as an example of surrogates, are very effective in estimating any complicated model. ANNs use a set of training data to learn the mapping process. Training data are a set of input parameters and their corresponding output values. Generating training data requires conducting numerous regional scale ground‐motion simulations. These numerical simulations are computationally challenging. Therefore, a step‐by‐step learning method should be employed to reduce the need for generating unnecessary training data. These methods are called active learning. In this study, we use a pool‐based query by committee (QBC) active learning method with effective initialization approach to study the performance of the models in the training process. We use a dataset that is generated for a moderate earthquake on a regional scale for anelastic attenuation studies with the focus on the estimation of peak ground velocity. The results show that active learning provides better performance in reducing generalization error than does passive learning while the same number of training data is used. Variation of performance with an increasing number of training data is significantly less in an active learning approach which indicates its stable and predictable behavior. This study, although limited to one earthquake and a metric, indicates that in developing surrogates as physics‐based earthquake ground‐motion simulators, application of active learning is an important step in reducing computational demands and generating stable predictors. 3. 标题：Pairwise Association of Seismic Arrivals with Convolutional Neural Networks. 作者：Ian W. McBrearty, Andrew A. Delorey, Paul A. Johnson. 日期：2019年1月9日 链接：https:///10.1785/0220180326 摘要导读： 正确地确定一个网络内陆震相位之间的关系对于发展准确的地震目录十分重要。几乎所有成熟方法用走时信息作为确定震相关系的主要准则。当地震活动性较高且许多假信号到达存在时，许多标准方法有可能无法准确解决问题。在这个工作中，我们利用卷积神经网络（CNNs）来解决震相相关性；我们训练CNNs读取一对儿台站上的地震波形到达信号，并利用二元分类法（Binary classification）确定两个波形是否来自同一源。将该方法应用到一个智利地震目录组成的大训练集，我们针对高频数据（>2Hz）获得了大于80%的真正确率（True possitive），并且台站间距超过100公里。额外的收获还包括，神经网络输出的结果还可以被用来预测到达信号的震相种类。该方法可以理想地与标准基于走时的相关方法结合，并且可以适应任意网络几何形状和应用，只要有足够多的训练数据。 Abstract: Correctly determining the association of seismic phases across a network is crucial for developing accurate earthquake catalogs. Nearly all established methods use travel‐time information as the main criterion for determining associations, and in problems in which earthquake rates are high and many false arrivals are present, many standard techniques may fail to resolve the problem accurately. As an alternative approach, in this work we apply convolutional neural networks (CNNs) to the problem of associations; we train CNNs to read earthquake waveform arrival pairs between two stations and predict the binary classification of whether the two waveforms are from a common source or different sources. Applying the method to a large training dataset of previously cataloged earthquakes in Chile, we obtain >80% true positive prediction rates for high‐frequency data (>2 Hz) and stations separated in excess of 100 km. As a secondary benefit, the output of the neural network can also be used to infer predicted phase types of arrivals. The method is ideally applied in conjunction with standard travel‐time‐based association routines and can be adapted for arbitrary network geometries and applications, so long as sufficient training data are available. 4. 标题：Convolutional Neural Network for Seismic Phase Classification, Performance Demonstration over a Local Seismic Network. 作者：Jack Woollam, Andreas Rietbrock, Angel Bueno, Silvio De Angelis. 日期：2019年1月16日 链接：https:///10.1785/0220180312 摘要导读： 在过去的二十几年里，可用的地震数据量显著地增加，加强了利用自动处理获取这些数据中包含的大量信息的需求。监测短期余震台网中的地震活动性是一个很有挑战的重要例子。这是因为这样的台网往往有着密集的台站覆盖，并且研究区域一般有较高的地震活动性。此外，如果需要区分震后不同的断层滑动物理过程，我们必须要得到非常精确地震定位；这就要求测出的地震波到时必须非常准确。这里，我们展示了一个使用卷积神经网络（Convolutional neural network）来为区域地震台网数据挑选震相的应用。CNN在一个小型数据集进行深度学习，基于411个在北部智利监测到事件（典型的余震流动台网）。虽然缺少大量数据，我们证明基于CNN的自动震相拾取仍旧能够提高震相分类的表现。其效果能够匹敌甚至超过之前的方法。训练好的神经网络被用来与Rietbrock et al.(2012)的优化的基于中期平均/长期平均方法（STA/LTA）进行比较，去分类另外一个有3878个事件的同一区域的数据集的震相。根据台站到时残差，CNN要比STA/LTA方法表现要好，并且达到地点残差分布接近于人工检测。 Abstract: Over the past two decades, the amount of available seismic data has increased significantly, fueling the need for automatic processing to use the vast amount of information contained in such data sets. Detecting seismicity in temporary aftershock networks is one important example that has become a huge challenge because of the high seismicity rate and dense station coverage. Additionally, the need for highly accurate earthquake locations to distinguish between different competing physical processes during the postseismic period demands even more accurate arrival‐time estimates of seismic phase. Here, we present a convolutional neural network (CNN) for classifying seismic phase onsets for local seismic networks. The CNN is trained on a small dataset for deep‐learning purposes (411 events) detected throughout northern Chile, typical for a temporary aftershock network. In the absence of extensive training data, we demonstrate that a CNN‐based automatic phase picker can still improve performance in classifying seismic phases, which matches or exceeds that of historic methods. The trained network is tested against an optimized short‐term average/long‐term average (STA/LTA) based method (Rietbrock et al., 2012) in classifying phase onsets for a separate dataset of 3878 events throughout the same region. Based on station travel‐time residuals, the CNN outperforms the STA/LTA approach and achieves location residual distribution close to the ones obtained by manual inspection. 5. 标题：Seismic Event and Phase Detection Using Time–Frequency Representation and Convolutional Neural Networks. 作者：Ramin M.H. Dokht, Honn Kao, Ryan Visser, Brindley Smith. 日期：2019年1月16日 链接：https:///10.1785/0220180308 摘要导读： 大量数字化地震波记录和深度学习在图像特征识别和分类问题中的成功给了我们一个可靠的地震检测方法的框架。传统方法的局限和挑战主要是不完备的模板波形（template waveforms）和低信噪比。我们利用深度卷积网络（deep convolutional network）设计了一个一般化的模型来提高地震和噪声信号的区别。利用西加拿大在3年内记录到的4900个地震作为数据集，一个多层深度卷积网络ConvNet被训练来识别背景噪声和地震信号在时间-频率域的特征。随后，我们训练了第二个网络利用主要地震信号到达的小波变换来区分P和S波，并估计它们的到时。验证试验表明两个网络都能打到近99%的准确率。我们将训练过的网络应用到一个近期在British Columbia东北记录的独立数据集。基于深度学习的方法好于传统方法。他们能够利用更少计算资源的情况下检测更多的地震事件。 Abstract: The availability of abundant digital seismic records and successful application of deep learning in pattern recognition and classification problems enable us to achieve a reliable earthquake detection framework. To overcome the limitations and challenges of conventional methods, which are mainly due to an incomplete set of template waveforms and low signal‐to‐noise ratio, we design a generalized model to improve discrimination between earthquake and noise recordings using a deep convolutional network (ConvNet). Exclusively based on a dataset of over 4900 earthquakes recorded over a period of 3 yrs in western Canada, a multilayer ConvNet is trained to learn general characteristics of background noise and earthquake signals in the time–frequency domain. In the next step, we train a secondary network using the wavelet transform of the major seismic arrivals to separate P from S waves and estimate their approximate arrival times. The results of validation experiments demonstrate promising performance and achieve an average accuracy of nearly 99% for both networks. To investigate the applicability of our algorithm, we apply the trained model on an independent dataset recently recorded in northeastern British Columbia (NE BC). It is found that deep‐learning‐based methods are superior to traditional techniques in detecting a higher number of seismic events at significantly less computational cost. 图片来自论文。 编辑校对：刘敦宇，吴葆宁。 ---------------------------------------------------------- 2019年伊始，我们的公众号 WhatAnEarthquake将开辟《地震前沿》版块，为地震科研工作者提供最新的科研论文导读。欢迎大家订阅！ |
|
|
来自： LibraryPKU > 《Geophysics》
