Abstract: Earthquake detection remains one of the fundamental operations in observational seismology. Despite the explosion in quality and quantity of seismic data our ability to build dense and complete seismicity catalogs remains limited. Recent approaches exploit the self-similarity of earthquakes and rely on using the waveforms of a few well known events as templates. Terabytes of continuous seismic signals at multiple stations are cross-correlated against these templates and thousands of events with high waveform similarity may be easily identified. This procedure, known as template matching, is computationally expensive and fails for events whose waveforms are significantly different from the templates. Also, the magnitudes of completeness of intermediate-depth earthquake catalogs are typically large. A dense catalog with a small magnitude of completeness not only allows for a detailed study of seismicity patterns, but also enables the mapping of small-scale b-value anomalies. We propose a Convolutional Neural Network (CNN) architecture for the simultaneous detection of intermediate-depth earthquakes, the precise picking of p- and s- wave arrival times and the estimation of their local magnitude. We test our implementation using a synthetically generated dataset and propose a training scheme that leverages both real and synthetic data for optimal results. As an example we apply our technique to a cluster of intermediate-depth intraplate earthquakes in northern Chile. We are able to detect eight times more events than in the initial catalog and use the picked arrival times to relocate them. We clearly resolve a double-planed structure at depth. Furthermore, we build a detailed b-value map and find that high b-value anomalies correlate well with regions were dehydration is expected.
Slidecast:
https://vimeo.com/276978538
Abstract: Specifying the location and extent of rifted crystalline crust (Precambrian basement) associated with the opening of the Iapetus Ocean in the eastern United States is important for seismic hazard evaluation. As currently mapped by the Central and Eastern U.S. Seismic Source Characterization for Nuclear Facilities (CEUS-SSC) Project, the western limit of major extensional faults and thinned extended crust produced by Iapetan rifting is within the Appalachian Plateau, to the west of the Valley and Ridge province. This limit was defined more than 20 years ago when little information was available about the crustal structure beneath the Appalachians and was based primarily on the distribution of earthquakes in two intraplate seismic zones, as well as some known basement faults. New estimates, using EarthScope, USArray and other data, show that the crustal thickness below the Blue Ridge, Valley and Ridge, and Appalachian Plateau provinces generally exceeds 45 km and is as much as 60 km in places. Such thick crust (equal to that typical of full-thickness continental crust) does not indicate significant extension as in rifted crust. We suggest that the western limit of major extensional faulting along the Iapetan rifted margin is located in the central Piedmont east of the Blue Ridge beneath the Blue Ridge–Piedmont megathrust sheet, and is closely associated with a prominent Bouguer gravity gradient (Appalachian gravity gradient). Only small Iapetan grabens and half grabens have been imaged with crustal seismic reflection data in the footwall beneath the Blue Ridge sheet. Our suggested location for the western limit of Iapetan rifting agrees with palinspastic reconstructions of the Iapetan rifted margin. Intracratonic grabens, inboard from the rifted margin, show only minor crustal extension and thinning. This seismotectonic model is a more credible representation of the basement below the southern and central Appalachians than the current model involving rifted crust.
Slidecast:
https://vimeo.com/276981987
Abstract: Earthquake ground motion records reveal period-dependent correlations, which has implication for seismic risk (Bayless and Abrahamson, 2017). The empirical inter-period correlations of epsilon (the residual between simulations and the mean of the simulations in Fourier Amplitude Spectra (FAS) space) using the Effective Amplitude Spectrum (EAS) computed from the PEER NGA-West2 database resemble a two-sided exponential function. We attempt to incorporate such correlation into the current San Diego State University (SDSU) Broadband (BB) ground motion generator module, which combines deterministic (low-frequency) and stochastic (high-frequency) components. Here, we assume that the Fourier amplitude at frequency f0 is correlated with the Fourier amplitude at f with correlation coefficient exp(-|f-f0|/a) and define a one-sided decaying exponential filter function g = H(f)exp(-f/a), where a is a constant. To improve the EAS correlation in the current SDSU module, we first generate uncorrelated uniformly distributed Fourier spectral amplitudes with unit mean for different realizations, and convolve them with g, which are then multiplied with the Fourier amplitude of the high-frequency ground motion synthetics calculated using Zeng et al. (1991)’s scattering theory. Using our improved method, the BB results for 7 western U.S. events and 2 Japan events with Mw5.0-7.2 show that the empirical inter-period correlations of EAS are well predicted in the SDSU module for a large number of realizations from a single event with unbiased goodness-of-fit of the spectral accelerations in the presence of correlated synthetics.
Slidecast:
https://vimeo.com/276979821
Abstract: The Caribbean-South American (CAR-SA) plate boundary is a complex transform fault system connecting oppositely vergent subduction zones, the Antilles in the east, and a currently locked CAR-SA flat slab subduction zone in the west. Teleseismic P-wave tomography shows both the Atlantic (ATL) and the Caribbean (CAR) plates subducting in opposite directions to transition zone depths under northern South America. Receiver functions (RF) show a depressed 660 discontinuity and thickened transition zone associated with each subducting plate. In the east, the ATL part of the SA plate subducts westward beneath the CAR. The eastern end of the El Pilar-San Sebastian strike-slip system, a subduction-transform edge propagator (STEP) fault, lies above the point where the ATL tears away from SA as it descends into the mantle. The Paria cluster seismicity is the mechanical expression of the plate tear. Body wave tomography and LAB depth determined from RFs and Rayleigh waves suggest that the descending plate also viscously removes the bottom third to half of the SA continental margin lithosphere. This has left thinned continental lithosphere under northern SA as the subduction zone has migrated along its northern coast. The thinned lithosphere extends almost the entire length of the El Pilar-San Sebastian fault system, from ~65o to 69oW, and inland more than 100 km. In northwestern SA the CAR plate subducts at < 30o to the ESE under northern Colombia to about Lake Maracaibo, Vn, and extends laterally from northernmost Colombia to perhaps as far south as the Bucaramanga nest seismicity. The flat slab is associated with three Neogene, basement cored, Laramide-style uplifts: the Santa Marta block, the Perija Range, and the Merida Andes. To the SE under Lake Maracaibo and the Merida Andes the CAR descends steeply to the transition zone. The steep descent suggests that the CAR plate is internally torn, separating the subducting CAR from CAR forming the seafloor to the north.
Slidecast:
https://vimeo.com/276984598
Abstract: The Probabilistic Seismic Hazard Maps for Japan (PSHMJ) have been criticized after the 2011 Tohoku earthquake. For example, Geller (2011) complained that recent damaging earthquakes occurred in places assigned a relatively low probability in the 2010 PSHMJ map of Intensity 6- or greater in 30 years. However, Miyazawa and Mori (2009, 2010) constructed maximum seismic intensity maps using recorded historical data for the past 500 years, and found that they agreed with the 2008 PSHMJ map of 10% in 50 years, which is equivalent to a hazard map of probable seismic intensity for a 475-year return period if we assume a Poisson distribution. We here explore a reason for the difference of the above two, using seismic intensities actually observed by the Japan Meteorological Agency and others. In Japan, we have seismic intensities actually observed in 1890 and later, and Miyakoshi et al. (2016) provided seismic intensity hazard curves at every grid point not only for the PSHMJ map of 2016 but also for those of 1890, 1920, 1950, and 1980. If we assume a Poisson distribution for seismic intensities like Miyazawa and Mori (2009), a PSHMJ map of 64% in 30 years is equivalent to a hazard map of probable seismic intensity for a 30-year return period. Therefore, we construct maximum seismic intensity maps using the observed intensities in 30 years from 1890, 1920, 1950, or 1980. We also construct PSHMJ maps of 1890, 1920, 1950, or 1980 using the intensities estimated for a probability of 64% in 30 years. We then compare the observed and estimated intensities at a same site, finding that the correlation between them is positive but very weak. This weakness may correspond to Geller’s (2011) criticism, probably because a period of 30 years is too short for a large intensity to be observed. If we adopt a period of 100 years, the correlation is certainly improved, approaching to the result of Miyazawa and Mori (2009) for a period of 500 years.
Slidecast:
https://vimeo.com/276982395
Abstract: Oklahoma has experienced an abrupt increase of induced seismicity in the last decade. Detection of small-magnitude earthquakes is essential because a complete earthquake catalog is the basis for understanding injection-induced seismicity in Oklahoma. Machine learning, especially deep learning, provides robust tools for image classification and feature extraction with complex structures. Convolutional neural networks (CNN) have been recently applied to continuous seismic waveform recordings to perform efficient phase picking and event detection (Zhu et al. 2017). In this work, we would like to verify the idea of transfer learning, which refines an existing classifier trained on a large dataset to a small dataset with a limited number of labeled samples in a different geographic region. Specifically, we transfer the knowledge learned from our recently developed CNN trained on a seismic dataset in China after the 2008 M7.9 Wenchuan earthquake (Zhu et al. 2017) to Oklahoma region. Using the single station CNN trained on the Wenchuan dataset, together with 895 local/regional catalog events recorded in central Oklahoma (Chen et al., 2018), we will refine part of the networks to pick the arrival times of the local seismicity in Oklahoma. The refined CNN results will then be compared with the matched filter results using the same catalog events as templates to verify its effectiveness. Updated results will be presented at the meeting.
Slidecast:
https://vimeo.com/276977865
