Seismometer Data Research Articles

Seismic evidence of an extremely thin crust along a 70 Ma slow-spreading crustal segment in the equatorial Atlantic Ocean

The oceanic crust at slow-spreading ridges is formed by a combination of magmatic and tectonic processes. Seismic tomographic studies, commonly used to determine crustal structures, indicate that mature slow-spreading crust has an average thickness of 6.1 ± 0.9 km, with a lower to upper crustal thickness ratio of 2.3. However, tomographic methods are based on high-frequency approximations and can only provide information about large-scale crustal structures. Using seismic full-waveform inversion applied to ocean bottom seismometer data, we report the presence of nearly uniform and extremely thin crust (4 ± 0.2 km) with a lower to upper crustal thickness ratio of 0.74 along an ∼100 km, 70 Ma crustal segment formed at the slow-spreading Mid-Atlantic Ridge in the equatorial Atlantic Ocean. This thin crust is underlain by a Moho transition zone that is 1 ± 0.5 km thick, resulting in a combined crustal and Moho transition zone thickness of 5 ± 0.6 km. The crustal thinning is primarily due to the thinning of the lower crust, suggesting that a significant part of the crust is formed by lava flows and dike intrusions. The presence of an extremely thin crust indicates a decrease in mantle temperature by at least 50 °C, suggesting a vast extent of the upper mantle thermal minimum in the equatorial Atlantic Ocean, and the absence of influence of the mantle from the Siera Leone mantle plume 70 m.y. ago.

Seismic Evidence for Velocity Heterogeneity Along ∼40 Ma Old Oceanic Crustal Segment Formed at the Slow‐Spreading Equatorial Mid‐Atlantic Ridge From Full Waveform Inversion of Ocean Bottom Seismometer Data

AbstractIn slow spreading environments, oceanic crust is formed by a combination of magmatic and tectonic processes. Using full waveform inversion applied to active‐source ocean bottom seismometer data, we reveal the presence of a strong lateral variability in the 40–48 Ma old oceanic crust formed at the slow‐spreading Mid‐Atlantic Ridge in the equatorial Atlantic Ocean. Over a 120 km‐long section between the St Paul fracture zone (FZ) and the Romanche transform fault (TF), we observe four distinct 20–30 km long crustal segments. The segment affected by the St Paul FZ consists of three layers, an ∼2 km thick layer with a P‐wave velocity <6 km/s, a 1.5 km thick middle crust with a velocity of 6–6.5 km/s, and an underlying layer where velocity is ∼7 km/s, representing the lower crust. The segment associated with an abyssal hill morphology contains a high velocity of ∼7 km/s at 2–2.5 km below the basement, indicating the presence of primitive gabbro or serpenized peridotite. The segment associated with a low basement morphology seems to have 5.5–6.5 km/s velocity starting near the basement extending down to ∼4 km depth, indicating chemically distinct crust. The segment close to the Romanche TF, a velocity 4.5–5 km/s near the seafloor increasing to 7 km/s at 4 km depth indicates a magmatic origin. The four distinct crustal segments have a good correlation with the overlying seafloor morphology. These observed strong crustal heterogeneities could result from alternate tectonic and magmatic processes along the ridge axis, possibly modulated by thermal and/or chemical variations in the mantle during their formation along the ridge segment.

A Steep Slab at the Yap Trench Resulted from Subducting Oceanic Plateaus

Abstract Interaction of oceanic plateaus with trenches plays a vital role in subduction activities and tectonic evolutions. The Yap trench is a rare case of an oceanic plateau subduction system. However, the knowledge of the impacts of plateau-trench interaction on subduction activity is still insufficient, due to a lack of seismological observations. Using ocean-bottom seismometer data near the Yap trench from April 2016 to May 2017, we conduct seismicity analyses in the Yap subduction zone by utilizing a machine-learning algorithm and matched-filter detections. The pattern of seismicity in the Yap trench exhibits characteristics similar to typical active subduction zones. The seismicity delineates a steep subducting slab, which may have resulted from the blocking of the buoyant Caroline Plateau. The majority of earthquakes are shallower than 80 km in the event-detectable area in the Yap trench, much shallower than the potential slab depth of 350 km from the previous seismic tomography images.

Revisiting the OBS seafloor compliance signal removal with a stationarity and stacking-based approach: the BRUIT-FM toolbox

Summary This study focuses on improving the seafloor compliance noise removal method, which relies on estimates of the compliance transfer function frequency response (the deformation of the seafloor under long-period pressure waves). We first propose a new multi-scale deviation analysis of broadband ocean bottom seismometer data to evaluate stationarity properties that are key to the subsequent analysis. We then propose a new approach to removing the compliance noise from the vertical channel data, by stacking daily estimated transfer function frequency responses over a period of time. We also propose an automated transient event detection and data selection method based on a cross-correlation criterion. As an example, we apply the method to data from the Cascadia Initiative (network 7D2011). We find that the spectral extent of long-period forcing waves varies significantly over time so that standard daily transfer function calculation techniques poorly estimate the transfer function frequency response at the lowest frequencies, resulting in poor denoising performance. The proposed method more accurately removes noise at these lower frequencies, especially where coherence is low, reducing the mean deviation of the signal in our test case by 27 % or more. We also show that our calculated transfer functions can be transfered across time periods. The method should allow better estimates of seafloor compliance and help to remove compliance noise at stations with low pressure-acceleration coherence. Our results can be reproduced using the BRUIT-FM Python toolbox, available at https://gitlab.ifremer.fr/anr-bruitfm/bruit-fm-toolbox.

Seafloor Rayleigh ellipticity, measured from unoriented data, and its significance for passive seismic imaging in the ocean

SUMMARY Observations of the seafloor Rayleigh ellipticity contribute to seismic imaging in the ocean. To extract such observables from the arbitrarily oriented ocean–bottom seismometer (OBS) data, we develop an orthogonal-regression-based approach to measure the waveform amplitude ratios of the unoriented horizontal and vertical components. The amplitude ratios are then used to calculate the Rayleigh ellipticity (and the sensor orientation angle). The robustness of our method is verified by applications to both the unoriented OBS data and the well oriented on-land seismic data. As we propose to calculate the Rayleigh ellipticity directly from the unoriented three-component data, the measurement process avoids the complexity arising from the surface wave non-great-circle effects and uncertainties of the OBS sensor orientation angles. Overall the Rayleigh ellipticity measurements from our method are systematically higher than those by conventional analysis and show less uncertainties. Our analyses suggest that the Rayleigh ellipticity curve (14–60 s), which could be retrieved from the raw broad-band OBS data, is effective to constrain the oceanic lithosphere structure, and the accurate measurement of Rayleigh ellipticity curve is important. The potential of seafloor Rayleigh ellipticity for seismic imaging in the ocean is evidenced by a case study of the Japan Basin, the Sea of Japan. Considering the insufficient station coverage in the ocean, the single-station measurement of seafloor Rayleigh ellipticity is of significance for OBS community.

Full waveform inversion of ocean bottom seismometer data from the oceanic Pacific Plate of the Japan trench

SUMMARY The geometrical and seismic structure of the Pacific oceanic Plate of the Japan Trench is essential for the understanding of earthquake activity in the area. Ocean Bottom Seismometer (OBS) data can be used, via ray-based tomography, to obtain estimates of properties such as crust thickness and structure, hydration and depth to the Moho boundary. The spatial resolution of these properties can be substantially improved by using the full waveform inversion (FWI) method. Most OBS data in this area are acquired with a sparse receiver spacing of 5–6 km, whereas FWI is assumed to work best with denser (1–2 km) receiver spacing. We show that FWI can be adapted to sparsely sampled data with better resolution than traveltime tomography. Using a 500 km long OBS longitudinal profile from the Japan Trench we obtain a detailed velocity structure of the crust, a better definition of the Moho boundary, a well-defined low-velocity layer in the lower crust and a clear spatial definition of areas with velocity inversions.

Downward continued ocean bottom seismometer data show continued hydrothermal evolution of mature oceanic upper crust

Abstract Heat flow measurements indicate hydrothermal activity in oceanic crust continues at least for 65 m.y. after formation. Hydrothermal activity progressively fills cracks and pores with alteration products, which is expected to lead to a trend of increasing seismic velocities with age. Compilations of seismic-P-wave velocity models inverted from ocean bottom seismometer (OBS) data have failed to detect such an aging trend beyond crustal ages of ca. 10 Ma. However, in these models, the velocities of the uppermost crust, where fluid flow would be most concentrated, are poorly resolved. This is because as the oceanic crust matures, the first crustal arrivals on OBS records (which best resolve upper crustal velocities using tomographic inversion), become hidden in the coda of the water wave. This may lead to the masking of any aging trend in the seismic velocities. For the first time, we show how including downward continuation (DC) in the analysis of OBS data collected across 65 Ma seafloor significantly improves measurements of the P-wave velocities of the upper crust. Our new analysis reveals a highly heterogeneous upper crust, with ridge-parallel P-wave velocity variations of 25%, implying local porosity values that are up to double that of global averages. Our new results, combined with other most recent advanced seismic analyses, reveal that seismic velocities indeed evolve with age up to at least 70 Ma, confirming that hydrothermal activity continues in mature oceanic crust.

Accurate Trace-Cut and Phase Alignment of Active Ocean-Bottom Seismometer Data

Abstract Accurate positions of ocean-bottom seismometers (OBSs) on the seafloor are critical parameters and can only be obtained by inversion modeling of first-arrival travel times of overhead cross-line airgun shootings. With an increased sampling interval of ≤20 ms for long-term earthquake studies, apparent artifacts affect the phase alignment of first arrivals on the seismic sections of trace-cut airgun shots. Our analysis shows that these apparent misalignments are caused by timing inconsistencies and inaccuracies during the trace-cut, which are so-called rounding errors. To eliminate these rounding errors, a simple interpolation is used to resample traces. Further analysis shows the simple interpolation satisfactorily retains the original waveform. The improved timing accuracy significantly reduces the uncertainty of seafloor locations as shown by Hadal OBS data.

Seismic Monitoring of a Deep Geothermal Field in Munich (Germany) Using Borehole Distributed Acoustic Sensing.

Geothermal energy exploitation in urban areas necessitates robust real-time seismic monitoring for risk mitigation. While surface-based seismic networks are valuable, they are sensitive to anthropogenic noise. This study investigates the capabilities of borehole Distributed Acoustic Sensing (DAS) for local seismic monitoring of a geothermal field located in Munich, Germany. We leverage the operator's cloud infrastructure for DAS data management and processing. We introduce a comprehensive workflow for the automated processing of DAS data, including seismic event detection, onset time picking, and event characterization. The latter includes the determination of the event hypocenter, origin time, seismic moment, and stress drop. Waveform-based parameters are obtained after the automatic conversion of the DAS strain-rate to acceleration. We present the results of a 6-month monitoring period that demonstrates the capabilities of the proposed monitoring set-up, from the management of DAS data volumes to the establishment of an event catalog. The comparison of the results with seismometer data shows that the phase and amplitude of DAS data can be reliably used for seismic processing. This emphasizes the potential of improving seismic monitoring capabilities with hybrid networks, combining surface and downhole seismometers with borehole DAS. The inherent high-density array configuration of borehole DAS proves particularly advantageous in urban and operational environments. This study stresses that realistic prior knowledge of the seismic velocity model remains essential to prevent a large number of DAS sensing points from biasing results and interpretation. This study suggests the potential for a gradual extension of the network as geothermal exploitation progresses and new wells are equipped, owing to the scalability of the described monitoring system.

Seismic noise in South China Sea: High-quality time-frequency analysis from 0.01 to 125 Hz.

An efficient and precise time-frequency analysis method for real-time ocean bottom seismometer (RTOBS) data in the South China Sea (SCS) is presented. Overcoming the limitations of conventional methods, the method involves temporal segmentation, unique frequency octaves, and Fourier transforms to generate power spectral density (PSD) and probability density function profiles. The method demonstrates superior precision, computational efficiency, and full-bandwidth (0 to Nyquist) capability compared to traditional techniques, as validated through theoretical and empirical evaluations. Applied to SCS RTOBS data, it unveils temporal PSD variations, shedding light on underwater noise sources like earthquakes, offshore blasting, ship-induced disturbances, and tidal effects. Establishing background noise levels in the SCS supports noise source categorization and ocean environment monitoring. Furthermore, comparing onshore and offshore seismic stations advances interdisciplinary research, fostering a comprehensive understanding of acoustics and seismology in the region.

HD-TMA: A New Fast Template Matching Algorithm Implementation for Linear DAS Array Data and Its Optimization Strategies

Abstract Distributed acoustic sensing (DAS) technology, combined with existing telecom fiber-optic cable, has shown great potential in earthquake monitoring. The template matching algorithm (TMA) shows good detection capabilities but depends on heavy computational cost and diverse template events. We developed a program named HD-TMA (high-efficiency DAS template matching algorithm), which accelerates computation by 40 times on the central processing unit platform and 2 times on the graphic processing unit platform. For linear DAS array data, we introduced a fast arrival-picking algorithm based on the Hough transform to pick the time window of template waveform. The HD-TMA was successfully applied to the 2022 Ms 6.9 Menyuan earthquake aftershock sequence recorded by a DAS array, and the DAS data result was compared with a collocated short-period seismometer data’s result. Two optimization strategies were discussed based on this data set. (1) Using signal-to-noise ratio in choosing the location and aperture of the subarray and the time window of the template waveform. (2) Considering the decrease in template events’ marginal utility, we proposed applying a neural network to build a template event library, followed by the HD-TMA scanning. Such strategies can effectively reduce computational cost and improve detection capability.

An efficient elastic full‐waveform inversion of multiple parameters with ocean‐bottom seismometer data

AbstractDetailed knowledge of the subsurface elastic properties might provide much‐needed insight into the subduction‐zone structure and the estimation of reservoir parameters. Compressional and shear wave velocities can be inverted by elastic full‐waveform inversion using multicomponent ocean‐bottom seismometer data. However, this process is computationally intensive, requiring massive, repeated simulations scaling with the number of sources. Although source–receiver reciprocity can streamline elastic full‐waveform inversion, it cannot be applied by directly interchanging the positions of sources and ocean‐bottom seismometers. To reduce the computational cost, we develop a source–receiver reciprocal elastic full‐waveform inversion, in which the reciprocity in multicomponent data can be accomplished by transforming the data matching from the original geometry to a specific form under the reciprocal geometry. This approach significantly reduces computational costs from twice the number of sources, as seen in conventional elastic full‐waveform inversion, to a scale with the count of ocean‐bottom seismometers. The tests on synthetic data verify that the proposed reciprocal elastic full‐waveform inversion maintains accuracy while improving computational efficiency, and further application on the field data from East China Sea also showcases the efficiency of the proposed method, resulting in a speed‐up of 10 times.

Reliable Earthquake Source Parameters Using Distributed Acoustic Sensing Data Derived from Coda Envelopes

Abstract A challenge in fully using distributed acoustic sensing (DAS) data collected from fiber-optic sensors is correcting the signals to provide quantitative true ground motion. Such corrections require considering coupling and instrument response issues. In this study, we show via comparison with geophone and broadband seismometer data that we can use coda envelope calibration techniques to obtain absolute moment magnitudes and source spectra from DAS data. Here, we use DAS and nodal geophones deployed as part of a geothermal monitoring experiment at Brady Hot Springs, Nevada, and on a 20 km long dark fiber of the ESnet’s Dark Fiber Testbed–a U.S. Department of Energy user facility, in Sacramento, California. Several DAS line segments with colocated geophone stations were used to compare the amplitude variation using narrowband S-wave coda envelopes. The DAS coda envelope decay at each point showed remarkable similarity with coda envelopes from different events in each narrow frequency range examined. The coda envelopes are used to determine Mw magnitudes and source spectra from regional stations without any major scatter. Because coda waves arrive from a range of directions, the azimuthal sensitivity of DAS is somewhat ameliorated. We show that the openly available seismic coda calibration software toolkit can be used for straightforward and faster processing of large DAS datasets for source parameters and subsurface imaging.

Retrieval and precise phase-velocity estimation of Rayleigh waves by the spatial autocorrelation method between distributed acoustic sensing and seismometer data

Summary In distributed acoustic sensing (DAS), optical fibre is used as sensors, which enables us to observe strain over tens of kilometres at intervals of several metres. S-wave velocity (Vs) structures of shallow sediments of high resolution have been obtained from surface wave dispersion curves by applying seismic interferometry to DAS data both onshore and offshore. However, it is known that there is a disadvantage to DAS seismic interferometry. In addition to Rayleigh waves, Love waves are also included. Consequently, the accuracy of the estimated phase velocities for Rayleigh waves is reduced due to the contamination of Love waves. To address this shortcoming, we suggest a spatial autocorrelation (SPAC) method between DAS and the vertical component of seismometer data. The SPAC method is equivalent to seismic interferometry and is useful for obtaining phase velocity dispersion curves of surface waves from the cross-correlation functions (CCFs) between the records of two receivers. The CCFs obtained from a combination of DAS and vertical seismometer data should contain only Rayleigh waves because Love waves have no vertical component. CCFs between DAS and vertical seismometer data are therefore expected to give more accurate phase velocities of Rayleigh waves than CCFs with DAS data only. In this study, we first formulated analytical expressions of cross-spectra for DAS and three-component seismometer data because seismic observation is generally carried out using a three-component seismometer. A new SPAC method is presented in the form of analytical expressions. We showed that our formulation only includes Rayleigh and not Love waves in the cross-spectra with DAS and the vertical-component seismometer data. We applied our SPAC method to actual DAS and vertical seismometer data recorded on the seafloor. Then, we compared our new SPAC method for DAS and vertical seismometer data with a conventional SPAC method for only DAS data. The results reveal that our new SPAC method can estimate the phase velocities of Rayleigh waves more accurately than the conventional method. In addition, the analytical formulations of the cross-spectrum between DAS and three-component seismometer data, which we obtained in this study, are expected to be useful for the estimation of accurate three-dimensional structures in the future, although this is not available at the moment due to the lack of an applicable dataset.

Origin of a high-velocity layer: Insights from seismic reflection imaging (South China Sea)

Magmatism and crustal structural characteristics are well-known phenomena that support understanding the marginal sea's tectonic evolution. The South China Sea (SCS) has become a natural laboratory for researchers to investigate the evolution of the marginal sea because of its complex tectonic background and multiple phases of magmatic activities. However, due to limitations in resolution and depth of observations, the evolutionary process of the SCS is still controversial. We use the latest processed multi-channel seismic reflection profile Line AB to evaluate the sedimentary sequences, crustal tectonics, and the Cenozoic magmatism in the southern SCS. The gravity modeling results indicate a reduction in crustal thickness from 18 km in the proximal domain to 7 km in the Continent-Ocean Transition (COT). The COT is roughly 35 km wide, representing a rapid break up process of the SCS. A high-velocity layer (HVL) of about 3 km thickness was formed in the continental slope crust, interpreted as post-spreading magmatic underplating. In addition, four sedimentary sequences, post-spreading magmatic intrusion, and the possible magma conduits were identified in the profile. Combining with previous ocean bottom seismometer (OBS) data, multi-channel seismic data, fault system data, and petrologic observations, we indicate the deep mantle magmatic material upwelled to the crust and further intruded into sediments along the hyper-extended crust or detachment faults during the post-spreading stage. We contend that cooling and heat shrinking within the SCS oceanic lithosphere have the potential to initiate decompression melt deformation in the continental slope. In summary, we construct a model of the origin and migration mechanism of extensive post-spreading magmatism in the SCS. Our research reveals the Cenozoic evolution and the HVL of the southern SCS, and provides a possible genesis for the anomalously young magmatism observed at the continental margin.

Ambient Noise Interferometry Using Ocean Bottom Seismometer Data From Active Source Experiments Conducted in the Southernmost Mariana Trench

AbstractOcean bottom seismometers (OBSs) have been used to detect submarine structural and tectonic information for decades. According to signal source controllability, OBS data have generally been classified into active and passive source data categories. The former mainly focuses on the compressional wave (P‐wave) velocity inversion and always lacks valid information about the shear wave (S‐wave) velocity structure. While the latter provides structural information with limited resolution due to the aperture of the stations. Overcoming the barriers between processing these two data types will allow the reuse of a vast amount of data from active source experiments to explore the submarine S‐wave velocity structural properties. Here, we creatively applied ambient noise interferometry to invert the S‐wave velocity structure using data from active source OBS deployment conducted in the southernmost Mariana subduction zone, which had already been utilized to detect submarine P‐wave velocity structure. Considering the short time duration and relatively low quality of this type of data, a combined method of short‐segment cross‐correlation and selected time‐frequency domain phase‐weighted stacking was adopted to obtain stable cross‐correlation functions, which were subsequently used to invert S‐wave velocity structures. Compared to previous studies using different methods, our result sheds new light on the crust and upper mantle structure of the southernmost Mariana subduction zone. This method could be used to detect more information based on the reutilization of existing active source OBS data.

3D Local Earthquake Tomography of the Andaman–Nicobar Subduction Zone Using Ocean-Bottom Seismometer Data

ABSTRACT The Andaman–Nicobar (A–N) subduction zone is one of the most seismically active subduction zones of the world where the Indian plate subducts beneath the Burmese–Sunda plate. Imaging the subducting Indian plate (SIP) geometry in this region is important to understand the subduction process, earthquake genesis, and associated seismic hazards. Therefore, we imaged the SIP for the first time using local earthquake data recorded from a network of nine ocean-bottom seismometers and six surface seismic stations. We inverted 2819 P and 2171 S phases picked from 410 local earthquakes recorded between December 2013 and May 2014 to obtain the tomographic images in the A–N region. The images show high-VP and VP/VS anomalies linked to colder and thicker SIP in the A–N region. We also observed seismic signatures of strong structural heterogeneity all along the SIP. The low-velocity anomaly at 60–100 km depth beneath the Andaman back-arc spreading center indicates mantle upwelling. Likewise, low-VP anomalies beneath the active volcano Barren Island indicate production of arc magmas by slab dehydration and corner flow in the mantle wedge.

On the Effect of 3D Wave Propagation on 2D Regional‐Scale Velocity Model Building

AbstractActive seismic surveys are routinely employed by academia to study geological structure of the crust and upper mantle. Wavefields generated during these surveys are sampled at the receiver locations, but the wave‐paths traveled from a source to a sensor remains unknown. Although seismic acquisition layouts designed to investigate complex crustal‐scale environments are often two‐dimensional, the seismogram recorded at the receiver location represents information gathered along the three‐dimensional wavepaths that might offset from the 2D source/receiver profile along its transverse direction. This so‐called 3D‐effect distorts the results of 2D seismic imaging, which is unable to handle the out‐of‐plane propagation. Despite the numerous 2D seismic imaging case studies, the assessment of this issue is often overlooked. However, the problem exists ‐ especially for crustal‐scale profiles, where seismic energy propagates over distances of hundreds of kilometers and probes different crustal units. In this work we investigate the impact of 3D‐effect on the results of 2D velocity model building from the academic ocean‐bottom seismometer data. We show with polarization analysis how the 3D‐effect can manifest itself in the data domain. Using various scenarios of acquisition we evaluate the imprint of the out‐of‐plane propagation on the data and the results of full‐waveform inversion. We show that 2D velocity model building from the seismic profiles acquired in the complex geological setting can lead to wrong solution. Looking for the remedy to this issue we couple different configurations of acquisition geometries with 3D full‐waveform inversion that allow to handle the 3D effect and provide correct model reconstruction.

A Practical Approach to Automatic Earthquake Catalog Compilation in Local OBS Networks Using Deep-Learning and Network-Based Algorithms

Abstract In land-based seismology, modern automatic earthquake detection and phase picking algorithms have already proven to outperform classic approaches, resulting in more complete catalogs when only taking a fraction of the time needed for classic methods. For marine-based seismology, similar advances have not been made yet. For ocean-bottom seismometer (OBS) data, additional challenges arise, such as a lower signal-to-noise ratio and fewer labeled data sets available for training deep-learning models. However, the performance of available deep-learning models has not yet been extensively tested on marine-based data sets. Here, we apply three different modern event detection and phase picking approaches to an ∼12 month local OBS data set and compare the resulting earthquake catalogs and location results. In addition, we evaluate their performance by comparing different subcatalogs of manually detected events and visually revised picks to their automatic counterparts. The results show that seismicity patterns from automatically compiled catalogs are comparable to a manually revised catalog after applying strict location quality control criteria. However, the number of such well-constrained events varies between the approaches and catalog completeness cannot be reliably determined. We find that PhaseNet is more suitable for local OBS networks compared with EQTransformer and propose a pick-independent event detection approach, such as Lassie, as the preferred choice for an initial event catalog compilation. Depending on the aim of the study, different schemes of manual repicking should be applied because the automatic picks are not yet reliable enough for developing a velocity model or interpreting small-scale seismicity patterns.

OBSTransformer: a deep-learning seismic phase picker for OBS data using automated labelling and transfer learning

SUMMARY Accurate seismic phase detection and onset picking are fundamental to seismological studies. Supervised deep-learning phase pickers have shown promise with excellent performance on land seismic data. Although it may be acceptable to apply them to Ocean Bottom Seismometer (OBS) data that are indispensable for studying ocean regions, they suffer from a significant performance drop. In this study, we develop a generalized transfer-learned OBS phase picker—OBSTransformer, based on automated labelling and transfer learning. First, we compile a comprehensive data set of catalogued earthquakes recorded by 423 OBSs from 11 temporary deployments worldwide. Through automated processes, we label the P and S phases of these earthquakes by analysing the consistency of at least three arrivals from four widely used machine learning pickers (EQTransformer, PhaseNet, Generalized Phase Detection and PickNet), as well as the Akaike Information Criterion (AIC) picker. This results in an inclusive OBS data set containing ∼36 000 earthquake samples. Subsequently, we use this data set for transfer learning and utilize a well-trained land machine learning model—EQTransformer as our base model. Moreover, we extract 25 000 OBS noise samples from the same OBS networks using the Kurtosis method, which are then used for model training alongside the labelled earthquake samples. Using three groups of test data sets at subglobal, regional and local scales, we demonstrate that OBSTransformer outperforms EQTransformer. Particularly, the P and S recall rates at large distances (>200 km) are increased by 68 and 76 per cent, respectively. Our extensive tests and comparisons demonstrate that OBSTransformer is less dependent on the detection/picking thresholds and is more robust to noise levels.