Characterization with dense array data of seismic sources in ...

SUMMARY

We examine the key sources identified in a collection of over 156,000 localizations obtained from a 26-day dataset gathered by a dense array deployed along the San Jacinto fault near Anza, California. The events were positioned using an array processing method known as Match Field Processing. Notably, the accuracy of event localization diminishes for those events occurring outside the array. Consequently, we differentiate events located within the array from those outside based on geometric criteria. By comparing the time distribution of these localizations to additional data, such as meteorological information and human activity, we pinpoint the nature of the predominant events identified using our methodology. We observe that most events situated outside the array are likely linked to surface structures influenced by wind. In contrast, some events under the array coincide with regional earthquakes and may represent diffractions caused by heterogeneities in the fault. The remaining events within the array are potentially originated from the fault itself.

For more information, please visit Dense Array Observations.

1 INTRODUCTION

Exploring the top 500 m of the Earth's crust, especially fault zones, poses significant challenges due to the extreme properties of shallow materials—such as very low P- and S-wave velocities (Vp and Vs), low attenuation Q values, and high Vp/Vs ratios. These factors lead to susceptibility to non-linear behavior and temporal changes. Local and regional seismic activities, along with meteorological influences like rain and temperature fluctuations, cause rock damage manifesting as reductions in Vp, Vs, and Q values. This damage predominately affects shallow regions near faults, with substantial coseismic changes in the top 100 m. To gain a detailed understanding of these fault zones, dense arrays have been temporarily positioned in recent years, facilitating high-resolution imaging and monitoring studies. The enhanced spatial coherence at high frequencies has enabled noise-based tomography to render finer images of faults like the Newport-Inglewood Fault and the San Jacinto Fault Zone, at scales ranging from hundreds to tens of meters.

The continuous recordings of dense arrays also enable the detection and localization of minor surface and subsurface events. Utilizing spatial coherence in dense seismic arrays, recent studies have shown that combining Match Field Processing (MFP) with Markov-Chain Monte Carlo (MCMC) sampling yields numerous new detections near the SJFZ. This approach enhances algorithm efficiency through an analytic homogeneous model dependent on only three parameters and significantly cuts down computational time compared to traditional grid search methods.

Continuous seismic waveforms mostly comprise signals related to wind, air/train/vehicle traffic, and other natural and artificial sources of ground motion. These signals resemble earthquake and tremor waveforms, leading to potential false detections of weak seismic events. To tackle this challenge, the MFP-MCMC localization method uses the analysis frequency as a 'filter' to differentiate between surface and depth sources: surface sources are detected at lower frequencies, whereas deeper sources are identified at higher frequencies. This method successfully located surface sources like shots and moving vehicles by determining the epicentral position and the apparent velocity of the propagating waves.

Shallow sources outside the array were also detected using the same methodology. However, using a homogeneous velocity model for source localization has its limitations due to the strong ambiguity between depth and velocity. Identifying the dominant sources outside the array is crucial for further analysis. In the past decade, machine learning algorithms trained with labeled datasets of common signals have significantly improved the classification of continuous waveforms and the detection of small events.

This paper aims to provide a comprehensive interpretation of the entire 26-day dataset for dominant shallow sources detected within or outside the dense array located in the damage structure of the SJFZ. The paper is structured as follows: In Section 2, we outline our event classification algorithm based on the MFP-MCMC outputs. Sections 3 and 4 offer a seismic interpretation of shallow events located either outside or beneath the array, respectively. Section 5 discusses the conducted research and potential future studies.

2 EVENT CLASSIFICATION METHODOLOGY

In our prior research, we analyzed a 26-day dataset recorded by a dense sensor array along the San Jacinto Fault. Our objectives included detecting very shallow events under poor SNR conditions, distinguishing between surface events caused by human activities and shallow microseismic events, and estimating their locations. We applied the Match Field Processing technique to achieve these estimations by comparing the data in successive time windows to a simple homogeneous acoustic forward model. This was performed in the frequency domain using the Bartlett operator, which projects the synthetic field onto the cross-spectral density matrix, capturing auto-correlation and inter-correlation between sensors.

Originally developed for simpler applications in the ocean acoustic field, the MFP technique computes the synthetic wavefield using a fixed medium velocity for a grid of point sources in three dimensions. Applying this method to near-surface source localization in seismic arrays near fault zones is more complex due to the medium's intricate velocity structure and the variety of noise sources contributing to recorded ground motion.

To address these challenges, we enhanced the method by using study frequencies to separate surface and depth sources. Using a center frequency of 4 and 16 Hz, we tailored sensitivity to surface and deeper sources, respectively. We also computed the synthetic wavefield for varying apparent velocities and source positions. By taking advantage of frequency filters and using apparent velocity as a depth proxy, we adopted a cylindrical wave approach for computational efficiency.

We calculated the probability density function (PDF) for each time window using the Metropolis–Hastings algorithm, assuming a uniform a priori probability distribution. The PDF is depicted as a cloud of points indicating the most probable epicentral position and the corresponding apparent velocity. We also included the Bartlett operator output, normalized between 0 and 1, using a color scale.

Evaluating the focal spot size was automated by fitting an ellipse to the envelope of the MCMC output at -3 dB, focusing only on points with Bartlett operator values greater than half the maximum. Events were categorized as inside or outside the array based on conditions involving short axis length, long axis length, and the location of output points relative to the array.

3 ANALYSIS OF EVENTS OUTSIDE THE ARRAY

To streamline analysis, localizations outside the array were reduced to directional vectors and velocities. We determined the backazimuth from the MCMC-MFP outputs by fitting a line to the PDF output and calculating the angle relative to north. Representing localizations in polar plots with apparent velocity as the radial coordinate, we observed dominant localizations around 309° with velocities ranging from 3500 to 5000 m/s, coinciding with the SJFZ's strike direction.

Hypothesizing that dominant sources outside the array are subsurface candidates, their apparent velocities suggested waves from depth but still traveling in the subsurface. The position could also indicate diving body waves emitted from surface sources. Using the 16 Hz frequency filters surface sources within the array, but distant P waves emitted by surface sources can be measured at such frequency.

Fig. 3(b) shows the analytical incidence angle associated with velocities ranging from 1000 to 6500 m/s. Deriving this incidence angle using Snell's law, we applied an average 1-D gradient velocity model from Mordret et al. The dominant source's apparent velocities correspond to incidence angles from 10° to 13°, indicating waves surfacing 2.2–3.9 km northwest of the array.

We investigated temporal variations in cumulative localizations to discern the source's nature. If anthropogenic, daily variations were expected. Fig. 5(a) shows localizations in 15-minute bins over the experimental period, with an average computed using a moving window of 4.5 hours.

Comparing daily localization variations to weather data showed a daily cycle in temperature and wind gust velocity, correlating with MCMC-MFP localization maxima at sunset and sunrise.

4 INSIDE-ARRAY LOCALIZATIONS

We identified around ∼6000 localizations under the array. Fig. 7(a) and (b) illustrate cumulative localizations in different apparent velocity ranges. Lower apparent velocities suggested shallower depths with a main cluster 200m from the geological fault trace, consistent with the velocity model revealing complex shallow structures.

We explored if we localized scatterers instead of sources by comparing our localization times to theoretical arrival times from a regional catalogue and borehole data. Regional earthquakes accounted for 50% of the localizations, identified through their timing relative to expected arrival times and visibility within borehole-recorded earthquake codas.

5 DISCUSSION AND IMPLICATIONS

This study underscores the abundance and diversity of seismic energy sources detectable in continuous waveforms. Localizing shallow sources in fault zones entails challenges due to weak energy, high frequency content, and poor SNR signals. Overlapping weak seismic events and remote seismicity, dominated by surface waves, further compound the challenge.

Dense array data yielded over 156,000 shallow depth localizations in 26 days, which provide insights into the rheology and dynamics of the top 500 m of the crust. While more than 90% of events were localized outside the array, likely originating from wind-triggered structures, around ∼6300 microseismic events were localized beneath the array.

Understanding the nature of these microseismic events requires further analysis. Integrating machine learning into the process could enhance classification and detection, leveraging complex conditions and expanded classes. Future studies could benefit from reprocessing data with adjustable time windows and adopting machine learning or template matching techniques.

ACKNOWLEDGEMENTS

ISTerre is part of Labex OSUG@2020. YBZ acknowledges support from the National Science Foundation (grant EAR-1818589) and the Department of Energy (award DE-SC0016520).

REFERENCES

Ben-Zion Y. et al., 2015. Basic data features and results from a spatially dense seismic array on the San Jacinto fault zone, Geophys. J. Int., 202(1), 370–380. 10.1093/gji/ggv142 Capon  J., 1969. High-resolution frequency-wavenumber spectrum analysis, Proc. IEEE, 57(8), 1408–1418. 10.1109/PROC.1969.7278 Chen Y., 2018. Automatic microseismic event picking via unsupervised machine learning, Geophys. J. Int., 212(1), 88–102. 10.1093/gji/ggx420 Chmiel M., Roux P., Bardainne T., 2019. High-sensitivity microseismic monitoring: automatic detection and localization of subsurface noise sources using matched-field processing and dense patch arrays, Geophysics , 84(6), .10.1190/geo2018-0537.1 Cros  E., Roux P., Vandemeulebrouck J., Kedar S., 2011. Locating hydrothermal acoustic sources at old faithful geyser using matched field processing, Geophys. J. Int. , 187(1), 385–393. 10.1111/j.1365-246X.2011.05147.x Diersen  S., Lee E.J., Spears D., Chen P., Wang  L., 2011. Classification of seismic windows using artificial neural networks, Proc. Comp. Sci., 4, 1572–1581. 10.1016/j.procs.2011.04.170 Gradon C., Moreau L., Roux  P., Ben-Zion Y., 2019. Analysis of surface and seismic sources in dense array data with match field processing and Markov chain Monte Carlo sampling, Geophys. J. Int., 218(2), 1044–1056. 10.1093/gji/ggz224 Hillers G., Roux P., Campillo M., Ben-Zion Y., 2016. Focal spot imaging based on zero lag cross-correlation amplitude fields: application to dense array data at the San Jacinto fault zone, J. geophys. Res., 121(11), 8048–8067. 10.1002/2016JB013014 Inbal A., Cristea-Platon T., Ampuero J.-.P., Hillers G., Agnew  D., Hough SE., 2018. Sources of long-range anthropogenic noise in Southern California and implications for tectonic tremor detection, Bull. seism. Soc. Am. 1–20., 10.1785/0120180130 Johnson  C.W.

74 0 0