Unsupervised Learning Used in Automatic Detection and Classification of Ambient‐Noise Recordings from a Large‐N Array

Cite this article as Chamarczuk, M., Y. Nishitsuji, M. Malinowski, and D. Draganov (2019). Unsupervised Learning Used in Automatic Detection and Classification of Ambient-Noise Recordings from a Large-N Array, Seismol. Res. Lett. 91, 370–389, doi: 10.1785/ 0220190063. We present a method for automatic detection and classification of seismic events from continuous ambient-noise (AN) recordings using an unsupervised machine-learning (ML) approach. We combine classic and recently developed array-processing techniques with ML enabling the use of unsupervised techniques in the routine processing of continuous data. We test our method on a dataset from a large-number (large-N) array, which was deployed over the Kylylahti undergroundmine (Finland), and show the potential to automatically process and cluster the volumes of AN data. Automatic sorting of detected events into different classes allows faster data analysis and facilitates the selection of desired parts of the wavefield for imaging (e.g., using seismic interferometry) and monitoring. First, using array-processing techniques, we obtain directivity, location, velocity, and frequency representations of AN data. Next, we transform these representations into vector-shapedmatrices. The transformeddata are input into a clustering algorithm (called k-means) to define groups of similar events, and optimizationmethods are used to obtain the optimal number of clusters (called elbow and silhouette tests). We use these techniques to obtain the optimal number of classes that characterize the AN recordings and consequently assign the proper class membership (cluster) to each data sample. For the Kylylahti AN, the unsupervised clustering produced 40 clusters. After visual inspection of events belonging to different clusters that were quality controlled by the silhouette method, we confirm the reliability of 10 clusters with a prediction accuracy higher than 90%. The obtained division into separate seismic-event classes proves the feasibility of the unsupervised ML approach to advance the automation of processing and the utilization of array AN data. Our workflow is very flexible and can be easily adapted for other input features and classification algorithms. Introduction Arrays with ever-increasing station counts are fundamental in seismology (Rost and Thomas, 2002). Developments of the nodal technology to acquire seismic data by the oil and gas industry brought the concept of large-number (large-N) arrays to academia (Hand, 2014). Their applications include structural imaging, studies of seismicity, and monitoring (e.g., Lin et al., 2013; Ben-Zion et al., 2015; Quiros et al., 2015; Karplus and Schmandt, 2018). Large-N arrays are often combined with long recording times, which facilitates seismic ambient-noise (AN) recordings (Karplus and Schmandt, 2018). AN is generally defined as a complex wavefield composed of the superposition of signals from natural and anthropogenic sources that are not generated specifically for the purpose of a study. Here, we address the issue of characterizing the AN content by developing automatic detection and classification of the various seismic events in the recorded wavefield. Because AN is recorded and stored during every regular continuous acquisition campaign (in particular, using nodal systems; Dean et al., 2015), our methodology is a step forward to maximize the information from passive recordings. The key technique using AN is called seismic interferometry (SI; Schuster et al., 2004; Wapenaar and Fokkema, 2006; Draganov et al., 2007; Wapenaar et al., 2008; Schuster, 2009). SI allows the retrieval of virtual-source records by correlating noise recordings between pairs of receivers. SI is considered a cost-effective alternative for controlled-source operations, especially when terrain access is an issue. Successful applications 1. Institute of Geophysics PAS, Warsaw, Poland; 2. Faculty of Civil Engineering and Geosciences, Delft University of Technology, CN Delft, Netherlands; 3. Petro Summit E&P Corporation, Tokyo, Japan *Corresponding author: mchamarczuk@igf.edu.pl © Seismological Society of America 370 Seismological Research Letters www.srl-online.org • Volume 91 • Number 1 • January 2020 Downloaded from https://pubs.geoscienceworld.org/ssa/srl/article-pdf/91/1/370/4910542/srl-2019063.1.pdf by Instytut Nauk-Geologicznych user on 03 January 2020 of AN SI can provide the velocity and structural information at the exploration scale (Draganov et al., 2009), and at the crustal scale (Ruigrok et al., 2010). Imaging and monitoring of the shallow crust require highfrequency data, that is, high-frequency sources recorded with high sampling rate (Niu and Yamaoka, 2018). Moreover, dense geophone arrays deployed in areas of abundant noise activity such as operating mine sites or volcanoes enable unaliased spatial sampling of the noise-source distribution characteristic for such high-seismicity areas (Rost and Thomas, 2002). For such continuous event-rich AN recordings, the preferred processing approach should be automatic and require minimum human interaction (Hansen and Schmandt, 2015). At the same time, most conventional array-processing techniques require high signal coherency across the array, implying important constraints on the array geometry, spatial extent, and data quality (Almendros et al., 1999). Therefore, for an existing dataset, the main interest is to optimize the processing time (including tuning array-processing parameters), especially in cases where months of human work could be necessary. Detection and classification of seismic signals using machinelearning (ML) already has awell-established history (Dowla et al., 1990; Dysart and Pulli, 1990; Wang and Teng, 1995; Del Pezzo et al., 2003; Wiszniowski et al., 2014). These studies were a step toward developing effective ML techniques for distinguishing tremors and earthquakes (Nakano et al., 2019), geyser-eruption signals detection (Yuan et al., 2019), earthquake early warning (Li, Meier, et al., 2018; Kong et al., 2019), and many automatic approaches for accurate earthquake-parameter estimation (Böse et al., 2008; Meier et al., 2015; Cuéllar et al., 2018; Ochoa et al., 2018), including the almost separate branch of accurate phasepicking methods (Chen, 2018; Zhu and Beroza, 2018). From the previously mentioned applications of seismic arrays, imaging studies (Araya-Polo et al., 2018), monitoring volcanic tremors (Malfante et al., 2018), and earthquake early warning (Kong et al., 2016; Li, Meier, et al., 2018) are a few of the most interesting and challenging areas for employing ML techniques. An inherent processing part related to some of the ML applications in seismology is the detection and classification of specific event types (Rouet-Leduc et al., 2017; Zhou et al., 2019), for example, signals (tremors) in volcano monitoring (Bhatti et al., 2016), specific precursors observed in the seismograms in earthquake early warning (Minson et al., 2018), and bodyor surface-wave events for AN SI imaging studies (Nakata et al., 2016). For some of these signals such as the volcanic tremors or body-wave events, the sole detection might not suffice because the waveforms of the different event types are similar to each other in the time domain. They might be differentiated only by minor features when transformed using signal representations (i.e., transformed to other domains) such as Fourier transform, envelope, autocorrelation, and kurtosis to name just a few from the long list of signal transformations used in the ML detection context (see the review in Malfante et al., 2018). These applications demand more detailed classification and assessment of AN event types performed in real time. Such processing appears to be a good target for combined ML and array-processing techniques. A hybrid approach that combines ML and array processing is an emphasized and anticipated development in seismology (Kong et al., 2018; Bergen et al., 2019). Motivation Here, we focus on the performance of ML for detecting and assessing different categories of seismic events present in continuous AN recordings. We rely on the fact that the appearance of AN in seismic records differs in amplitude and frequency and that the various kinds of signal transformations derived therefrom provide varying characteristics (Bormann, 1998). We aim for evaluation of the feasibility of unsupervised clustering methods using these differences to track down various AN events without the need to know their exact representation in different domains. These characteristics can be retrieved using arrayprocessing techniques (Rost and Thomas, 2002). This idea is applied to data recorded using a large-N array above the Kylylahti active undergroundmine in eastern Finland for testing AN SI for mineral exploration. The same dataset has already been used to demonstrate extraction of body-wave events using supervised ML (support-vector machine) in combination with a two-step wavefield evaluation and detection method (Chamarczuk et al., 2019). With this hybrid approach, the authors made a binary classification to discriminate between bodyand surface-wave events. They applied two-step wavefield evaluation and detection method on a small portion of the AN recordings to obtain labels (in this study, label is the type of recorded seismic event) and then used the labels as input to a support-vector machine to classify the remaining part of the data. In our study, we do not provide input labels and aim to discriminate between several (>2) classes of seismic events and thus provide a more detailed description of the recorded AN data. In addition to the lack of labels, typical for unsupervised clustering, we treat the number of clusters as an unknown parameter to be established (we refer to this approach as “blind clustering”). In blind clustering, we estimate the range of optimal number of clusters by comparing the results of k-means for a broad range o