Ambient noise seismic interferometry (ANSI) extracts waves travelling between two stations by cross-correlating the time series data recorded by each station and thus estimates an Empirical Green’s Function (EGF), the impulse response of the Earth’s subsurface between the stations. One station acts as a “virtual source” and the other as a “receiver” (Dantas et al., 2018; Draganov et al., 2013; Nakata et al., 2015; Olivier et al., 2015; Panea et al., 2014; Poli et al., 2012; Quiros et al., 2016; Shapiro & Campillo, 2004; Thangraj et al., 2020a; Thangraj & Pulliam, 2021; Wang et al., 2014; Wapenaar et al., 2010, 2014). However, for virtual source gathers to be an accurate estimate of the EGF there must be 1) a dense and homogeneous distribution of uncorrelated noise sources and 2) an equipartitioned noise wavefield (Wapenaar et al., 2005). Unfortunately, these conditions are rarely met in real applications. To compensate for a non-uniform distribution of sources, correlations computed for short time windows (e.g., several minutes) are stacked over hours or days to produce a robust estimate of an EGF. To meet the second condition, sources are required to be located on the axis that connects the two stations. For the case of reflected waves, sources are required to be located in stationary phase zones (Draganov et al., 2013) or in a region in which a reflection between two receivers can be retrieved. In this study, cross-correlation results for a single time window is referred to as a single “cross-correlation panel”. Final EGF estimates result from stacking a set of cross-correlation panels.
Stacking correlations over a long period of time was proposed previously as a way to attain the broadest possible distribution of sources (Schimmel et al., 2011; Snieder, 2004) and increase the signal-to-noise ratio of arrivals by suppressing random noise. However, it has been challenging to establish how long correlations need to be stacked, or for passive seismic surveys to acquire data, to ensure that EGFs converge. The convergence rate of noise correlations has been studied by Sabra et al. (2005), who established theoretical relations between variance, recording length, frequency bandwidth and recorded energy. EGF convergence also depends on the choice of window length, processing parameters, and quality of data (Chamarczuk et al., 2021a; Medeiros et al., 2015; Weemstra et al., 2014).
Surface wave studies are more common than studies of body waves in ANSI (Agrawal et al., 2015; Bensen et al., 2007; Chmiel et al., 2019; Nishida et al., 2009; D. A. Quiros et al., 2018; Y. Yang et al., 2007) because surface waves dominate ambient noise sources (and recordings). Consequently, extracting body waves from ANSI is difficult and relatively few successful studies are found in the literature (Chamarczuk et al., 2021b; Cheraghi et al., 2017; Draganov et al., 2009; Panea et al., 2014; Roux et al., 2016; Ruigrok et al., 2011; Thangraj & Pulliam, 2021; Zhan et al., 2010). A particular challenge in retrieving body wave arrivals is the indeterminate nature of seismic sources, and insufficient knowledge of source locations, frequency characteristics and their variation with time. It can be shown that naively stacking all data in ANSI does not always result in the retrieval of body wave arrivals (Girard & Shragge, 2019; Olivier et al., 2015; Thangraj & Pulliam, 2021). Selectively stacking a subset of correlation panels in which body wave energy dominates can extract high signal-to-noise ratio body wave arrivals; selective stacking based on post-correlation metrics has been shown to be successful at extracting body waves (Cheraghi et al., 2017; Olivier et al., 2015; Panea et al., 2014; Safarkhani & Shirzad, 2019; Thangraj & Pulliam, 2021). With the exception of Thangraj & Pulliam (2021), the above-mentioned studies require an estimate of velocity or slowness to discriminate between the cross-correlation panels. There are few pre-correlation automated selective stacking methods (Draganov et al., 2013). Pre-correlation estimates can save significant computation time in computing cross-correlations and also would be less biased by pre-processing routines used in ANSI. In this study we show that it is possible to identify noise windows in which body wave energy dominates with the aid of machine learning methods, specifically unsupervised clustering.
Machine learning methods have been used extensively in seismology. Examples include earthquake phase picking (Soto & Schurr, 2021, Suarez & Given, 2021, Ming et al., 2019), seismic phase association (Ross et al., 2019), seismic wave discrimination (Li, Meier, et al., 2018), earthquake detection (Thomas et al., 2021; S. Yang et al., 2020; Walter et al., 2020), earthquake localization (Shen & Shen, 2021) and event discrimination (Renouard et al., 2021; Linville et al., 2019). However, most of the above-mentioned studies can be classified as “supervised learning” approaches because they require a training dataset and information about the data beforehand. Supervised techniques may therefore work well in the lab, when applied to previously-acquired data, but applications in the field, in real time, would be challenging or impossible. Ambient noise studies are often carried out on datasets for which little information is available beforehand, so “unsupervised clustering” or “unsupervised machine learning”, if feasible, would be better suited to exploring ambient noise datasets. Unsupervised clustering has been used previously to discriminate between local and teleseismic arrivals (Mousavi et al., 2019), to automate microseismic event picking (Chen, 2020), and to detect seismic events (Li, Peng, et al., 2018a). However, few studies have applied and assessed the capabilities of unsupervised learning methods to identify signals recorded by a passive seismic array. Johnson et al. (2020) and Chamarczuk et al. (2020) report on applications of unsupervised learning methods that succeed in identifying classes of signals in seismic data. Identifying noise windows with dominant body wave energy in ANSI studies remains an important and often time-consuming problem. This study assesses the prospects for unsupervised learning methods to save the time and effort required to separate these panels (Draganov et al., 2013; Vidal et al., 2014). We engineer features and tailor unsupervised learning to identify clusters in which body waves arrivals dominate cross-correlation panels, and ultimately selectively stack panels within clusters to reveal body wave arrivals with higher signal-to-noise ratios than is found by stacking all available panels. We also show that machine learning algorithms can help extract information required to label data for subsequent supervised learning studies.