Koper, Keith D.; Burlacu, Relu; Murray, Riley; Baker, Ben; Tibi, Rigobert; Mueen, Abdullah
Determining the depths of small crustal earthquakes is challenging in many regions of the world, because most seismic networks are too sparse to resolve trade-offs between depth and origin time with conventional arrival-time methods. Precise and accurate depth estimation is important, because it can help seismologists discriminate between earthquakes and explosions, which is relevant to monitoring nuclear test ban treaties and producing earthquake catalogs that are uncontaminated by mining blasts. Here, we examine the depth sensitivity of several physics-based waveform features for ∼8000 earthquakes in southern California that have well-resolved depths from arrival-time inversion. We focus on small earthquakes (2 < ML < 4) recorded at local distances (< 150 km), for which depth estimation is especially challenging. We find that differential magnitudes (Mw= ML–Mc) are positively correlated with focal depth, implying that coda wave excitation decreases with focal depth. We analyze a simple proxy for relative frequency content, Φ≡ log10 (M0)+3log10 (fc (,and find that source spectra are preferentially enriched in high frequencies, or “blue-shifted,” as focal depth increases. We also find that two spectral amplitude ratios Rg 0.5–2 Hz/Sg 0.5–8 Hz and Pg/Sg at 3–8 Hz decrease as focal depth increases. Using multilinear regression with these features as predictor variables, we develop models that can explain 11%–59% of the variance in depths within 10 subregions and 25% of the depth variance across southern California as a whole. We suggest that incorporating these features into a machine learning workflow could help resolve focal depths in regions that are poorly instrumented and lack large databases of well-located events. Some of the waveform features we evaluate in this study have previously been used as source discriminants, and our results imply that their effectiveness in discrimination is partially because explosions generally occur at shallower depths than earthquakes.
Seismic waveform data recorded at stations can be thought of as a superposition of the signal from a source of interest and noise from other sources. Frequency-based filtering methods for waveform denoising do not result in desired outcomes when the targeted signal and noise occupy similar frequency bands. Recently, denoising techniques based on deep-learning convolutional neural networks (CNNs), in which a recorded waveform is decomposed into signal and noise components, have led to improved results. These CNN methods, which use short-time Fourier transform representations of the time series, provide signal and noise masks for the input waveform. These masks are used to create denoised signal and designaled noise waveforms, respectively. However, advancements in the field of image denoising have shown the benefits of incorporating discrete wavelet transforms (DWTs) into CNN architectures to create multilevel wavelet CNN (MWCNN) models. The MWCNN model preserves the details of the input due to the good time–frequency localization of the DWT. Here, we use a data set of over 382,000 constructed seismograms recorded by the University of Utah Seismograph Stations network to compare the performance of CNN and MWCNN-based denoising models. Evaluation of both models on constructed test data shows that the MWCNN model outperforms the CNN model in the ability to recover the ground-truth signal component in terms of both waveform similarity and preservation of amplitude information. Model evaluation of real-world data shows that both the CNN and MWCNN models outperform standard band-pass filtering (BPF; average improvement in signal-to-noise ratio of 9.6 and 19.7 dB, respectively, with respect to BPF). Evaluation of continuous data suggests the MWCNN denoiser can improve both signal detection capabilities and phase arrival time estimates.
The Redmond Salt Mine (RSM) Monitoring Experiment in Utah was designed to record seis-moacoustic data at distances less than 50 km for algorithm testing and development. During the experiment from October 2017 to July 2019, six broadband seismic stations were operating at a time, with three of them having fixed locations for the duration, whereas the three other stations were moved to different locations every one-and-half to two-and-half months. RSM operations consist of nighttime underground blasting several times per week. The RSM is located in proximity to a belt of active seismicity, allowing direct comparison of natural and anthropogenic sources. Using the recorded data set, we built 1373 events with local magnitude (ML) of −2.4 and lower to 3.3. For 75 blasts (RMEs) from the Redmond Salt Mine and 206 tectonic earthquakes (EQs), both ML and the coda duration magnitude (MC) are well constrained. We used these events to test and design discriminants that separate the RMEs from the EQs and are effective at local distances. The discriminants consist of ML −MC, low-frequency Sg to high-frequency Sg, Pg/Sg phase-amplitude ratios, and Rg/Sg spectral amplitude ratios, as well as different combinations of two or more of these classifiers. The areas under the receiver operating characteristic curves (AUCs) of 0.92–1.0 for ML −MC, low-frequency Sg to high-frequency Sg, and Rg/Sg indicate that these discriminants are very effective. Conversely, the AUC of only 0.57 for Pg/Sg suggests that this discriminant is only slightly better than a random classifier. Among the effective classifiers, Rg/Sg, shows the lowest likelihood of misclassification (4.3%) for the populations. Results of joint discriminant analyses suggest that even the arguably inef-fective single classifier, like Pg/Sg in this case, can provide some value when used in combi-nation with others.
In this report we discuss training a deep learning seismic signal detection model on 3-component stations from the International Monitoring System (IMS) using the PhaseNet architecture. Using 14 years of associated signals from the International Data Centre’s (IDC) Late Event Bulletin (LEB), we auto-curated training data consisting of signal windows containing associated arrivals, and noise windows that contain no LEB-associated signals. We trained several models using different waveform window durations (30 seconds and 100 seconds), with and without bandpass filtering. We evaluated the effectiveness of our models using associated signals from the Unconstrained Global Event Bulletin (UGEB) and found that several of our models outperformed the signal detections from the IDC’s Selected Event List 3 (SEL3) arrival table. The SEL3 bulletin evaluated on the UGEB dataset with 100-second waveform windows registered a precision and recall of .15 and .48, respectively, versus .19 and .59 for our filtered-data model. For the 30-second waveform window dataset, the SEL3 bulletin achieved a precision and recall of .31 and .47, respectively, versus .32 and .60 for our filtered-data model. Finally, our models detected signals from all source-to-receiver distances, suggesting it is feasible to use a single PhaseNet model for the IMS network.
Waveform cross-correlation is a sensitive phase-matched filtering technique that can detect seismic events for nuclear explosion monitoring. However, there are outstanding challenges with correlation detectors, most notably a direct dependence on the completeness of the waveform template library. To ameliorate these challenges, we investigate how dynamic time warping (DTW) may make waveform correlation more robust. DTW analyzes the differences between two time series and attempts to “warp” one time series relative to another in a recursive manner. We apply DTW to synthetic earthquake and recorded explosion templates to expand the capability of correlation detectors. We explore what conditions (e.g., source, station distance, frequency bands) and/or DTW algorithms generate stronger correlation scores. We show that DTW performs well on noisy signals and can dramatically improve the cross-correlation coefficient between a template and data-stream waveform. We conclude with recommendations on how to utilize DTW in nuclear monitoring detection.
Aftershock sequences are a burden to real-time seismic monitoring. Cross-correlation can be used because aftershocks exhibit similar waveforms, but the method is computationally expensive. Deep learning may be an alternative, as it is computationally efficient, but great attention to training and testing is required in order to trust that the model can generalize to new aftershock sequences. This is problematic for aftershock sequences, because large-magnitude earthquakes are unpredictable and are globally widespread. Here, we test several paired neural network (PNN) models trained on a augmented (noise-added) earthquake dataset, to determine whether they can be generalized to process real aftershock sequences. Two aftershock datasets that were originally detected by cross-correlation and subsequently validated by an expert analyst were used. We found that current PNN models struggle to generalize to aftershock sequences. However, we identify approaches to improve training future PNN models and believe that improvements may be achieved by transfer learning.
Seismic waveform data are generally contaminated by noise from various sources, which interfere with the signals of interest. In this study, we implemented and applied several noise suppression methods using data recorded by the regional network of the University of Utah Seismograph stations. The denoising methods, consisting of approaches based on nonlinear thresholding of continuous wavelet transforms (CWTs, e.g., Langston and Mousavi, 2019), convolutional neural network (CNN) denoising (Tibi et al., 2021), and frequency filtering, were all subjected to the same analyses and level of scrutiny. We found that for frequency filtering, the improvement in signal-to-noise ratio (SNR) decreases quickly with decreasing SNR of the input waveform, and that below an input SNR of about 32 dB the improvement is relatively marginal and nearly constant. In contrast, the SNR gains are low at high-input SNR and increase with decreasing input SNR to reach the top of the plateaus corresponding to gains of about 18 and 23 dB, respectively, for CWT and CNN denoising. The low gains at high-input SNRs for these methods can be explained by the fact that for an input waveform with already high SNR (low noise), only very little improvement can be achieved by denoising, if at all. Results involving 4780 constructed waveforms suggest that in terms of degree of fidelity for the denoised waveforms with respect to the ground truth seismograms, CNN denoising outperforms both CWT denoising and frequency filtering. Onset time picking analyses by an experienced expert analyst suggest that CNN denoising allows more picks to be made com-pared with frequency filtering or CWT denoising and is on par with the expert analyst’s processing that follows current operational procedure. The CWT techniques are more likely to introduce artifacts that made the waveforms unusable.