SVR-DAS: A Machine-Learning-Based Method to Create Earthquake Catalogs from Seafloor Distributed Acoustic Sensing Measurements

SVR-DAS: A Machine-Learning-Based Method to Create Earthquake Catalogs from Seafloor Distributed Acoustic Sensing Measurements | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article SVR-DAS: A Machine-Learning-Based Method to Create Earthquake Catalogs from Seafloor Distributed Acoustic Sensing Measurements Gerardo Mendo-Pérez, Hiromichi Nagao, Shinya Katoh, Masanao Shinohara This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8080988/v2 This work is licensed under a CC BY 4.0 License Status: Posted Version 2 posted You are reading this latest preprint version Show more versions Abstract Seafloor earthquake monitoring is one of the emergent research areas in seismology. Its importance relies on the continuous monitoring of earthquake activity near or on subduction zones, where large megathrust earthquakes are generated. In Japan, seafloor observation systems have been deployed to monitor earthquake activity in Nankai area (DONET) and Japan Trench (S-net). However, these are limited to the number of observation points of the system, and their installation and maintenance can be logistically and economically expensive. Distributed Acoustic Sensing (DAS) measurements obtained from fiber optic cables are arising as a promising technology to monitor offshore earthquake activity. In addition, DAS data is suitable for deep learning methods due to the vast volume of data that it produces. Although successful methods have been developed for land-based DAS measurements, their performance in seafloor DAS data is compromised by additional sources of noise and coupling issues. State-of-the-art research is aimed at developing novel deep-learning-based methods to process seafloor data. However, DAS-based earthquake catalogs are still scarce, and manual picking is not feasible. In this work, we propose a novel machine-learning-based method to build earthquake catalogs from seafloor DAS measurements. Our method detects earthquakes by applying envelope template matching using a small number of templates obtained from DAS data itself. P and S arrival times are obtained from detections using STA/LTA and Autoregressive models using the Akaike Information Criterion (AR-AIC). The coincident picks of both methods are used as training data to create P and S picks models using Support Vector Regression (SVR). For this reason, our method is called SVR-DAS. We test the method using relatively high Signal-to-Noise Ratio (SNR) DAS data from an earthquake that occurred in 2023-07-26. Subsequently, we applied SVR-DAS to the DAS data from 2022-02-28, obtaining P and S models from relatively high SNR DAS data. The created catalogs are used to train and validate a CNN-RNN model. In addition, we present a comparison of the SVR-DAS picks with the picks obtained from PhaseNet-DAS, a deep-learning-based model for earthquake detection and seismic picking. Our results prove that SVR-DAS is capable to build earthquake catalogs from DAS data with enough accuracy. The obtained catalogs are useful for further development of deep learning models for seafloor DAS data processing. Seismology Artificial Intelligence and Machine Learning Distributed Acoustic Sensing Deep Learning Support Vector Regression earthquake detection seismic picking Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Introduction Monitoring offshore earthquake activity is an important research field for seismic hazard assessment. Most of the large megathrust source areas are located offshore near subduction zones. For this reason, marine earthquake monitoring systems have been installed over the last 20 years, allowing us to obtain offshore seismic and geodesic observations. In Japan, seafloor observatories such as the Dense Ocean Floor Network System (DONET) (Kawaguchi et al., 2011 ; Aoi et al., 2020 ), N-net (Aoi et al., 2023 )d net (Kanazawa et al., 2016 ; Aoi et al., 2020 ) are currently operating in Nankai and Japan Trench regions, respectively. Although several advances have been achieved by current marine infrastructure, establishing dense marine observation networks remains a challenge due to the difficulty to install seafloor observatories in deeper areas and recovering data. In addition, the maintenance and lifespan of seafloor seismometers is very limited, and this can result in extremely costly observation systems. An emergent solution to the observational gaps in deep marine areas is the use of Distributed Acoustic Sensing (DAS) measurements, which is transforming earthquake monitoring. DAS measures the phase changes between an induced laser pulse and Rayleigh backscattering produced by inhomogeneities in the fiber structure (Lindsey and Martin, 2020 ). This phase change is linearly proportional to the strain of fiber optic cables to earthquakes or any other external source that perturbs the cable. The interrogator unit (IU) sends controlled laser pulses at a specific wavelength (near 1550 nm) throughout the optical fiber, and a phase change of the scattered light is measured by the IU (Masoudi and Newson, 2016 ; Hartog, 2017 ; Lindsey and Martin, 2020 ; Zhan, 2020). The strain rate is proportional to the change in the gauge length of scattered light obtained by emitting laser pulses. Hence, DAS enables fiber optic cables to act as linear strainmeters (Benioff, 1935 ), allowing the retrieval of earthquake signals (Shinohara et al., 2019 ; 2021; Lindsey and Martin, 2020 ). Two main advantages arise from the use of seafloor optic fiber to monitor earthquakes. First, DAS measurements are equivalent to a linear array of thousands of channels, allowing to record earthquakes with unprecedented resolution (Lindsey and Martin, 2020 ). The second is that DAS measurements can be done using spare (dark) fibers from already installed telecommunication cables. DAS was first applied in oil and gas industry, where encased fiber optic cables in boreholes were used for seismic exploration (Daley et al. 2013 ). In seismology, teleseismic waves have been retrieved from broadband DAS measurements (Lindsey et al., 2017 ; Yu et al., 2019 ). Despite its promise, many challenges remain to be overcome. One is the manageability of high volumes of data. The size of one day of continuous DAS data may rise to several tens of TB (depending on the sampling rate and the gauge length). Consequently, large computing clusters are required for storage, which may be insufficient for real-time applications. Also, it is impossible for analysts to process high-density data in a reasonable time frame. Another issue is related to the cable coupling to the ground. Several laboratory experiments suggest that loosely installed cables produce weaker responses (Papp et al., 2017; Becker et al., 2018; Lindsey and Martin, 2020 ; Ajo-Franklin et al., 2021). In field observations, urban activity often produces stronger signals than earthquakes (Lindsey and Martin, 2020 ), and in colder areas, the sensitivity to high-frequency sources is variable in the presence of snow and ice (Castongia et al., 2017 ; Walter et al., 2020 ; Lindsey and Martin, 2020 ). In recent years, deep learning has emerged as a new paradigm in science for data processing due to its ability to make data-driven decisions based on nonlinear functions. In seismology, well-established methods have been developed for earthquake detection and phase picking (Ross et al., 2018 ; Zhu and Beroza, 2019 ; Mousavi et al., 2020 ; Kaneko et al., 2023 ; Tokuda and Nagao, 2023 ; Katoh et al., 2025 ), seismic phase association (Ross et al., 2019 ; Zhu et al., 2022 ; McBrearty and Beroza, 2023 ), polarity determination and focal mechanism estimation (Zhang et al., 2023 ; Katoh et al., 2025 ), denoising (Yu et al., 2019 ), waveform inversion (Liu et al., 2021 ), and modeling (Moseley et al., 2020 ). For DAS data, there are already proposed methods for earthquake detection and seismic picking yielding accurate results (Zhu et al., 2023 ; Ding et al., 2025 ). Current deep learning models used for DAS data are trained using inland seismic and DAS records. While these models successfully capture broad earthquake-related features related to the earthquake signal, finer details associated with noise present in seafloor environments are not being considered, leading to errors in seismic phase picking for some events. To the best of our knowledge, the seafloor DAS-based training datasets of seafloor optic fiber cables are still scarce and currently is an active research field (Xiao and Tillman, 2024). In this work, we address this obstacle by creating training/validation datasets from seafloor DAS data that can be used to develop and/or test deep learning models. Our framework consists of two stages: earthquake detection and body wave phase picking (P and S). In the detection stage, we apply template matching using envelope waveforms to identify earthquake signals in DAS data. We begin with one template retrieved from DAS data, and all subsequent detections are used to expand the template library. In the phase picking stage, we apply recursive STA/LTA (Allen, 1982 ) and autoregressive models using the Akaike Information Criterion (AR-AIC) (Akazawa et al., 2004) to obtain P and S picks. We compare the obtained picks, and the consistent ones are selected as training datasets to create P and S pick models. To create these models, we apply Support Vector Regression (SVR) (Vapnik, 1997 ), a machine-learning method that searches for hyperplanes that best fits the data. First, we test the method using well-defined DAS data from an earthquake that occurred near the cable on 2023-07-26 12:26:00 UTC. Afterwards, we apply SVR-DAS on DAS data from 2022-02-28. Finally, we train and validate a CNN-RNN model inspired by trigger-word systems (Supriya et al., 2020 ). The obtained picks from SVR-DAS are compared with picks from PhaseNet-DAS (Zhu et al., 2023 ), a deep-learning-based method for earthquake detection and seismic picking in DAS data. We believe that SVR-DAS method will provide accurate arrival-times estimates from seafloor DAS data. Data and Methods This work focuses on the detection of earthquake signals and the identification of P and S phases in seafloor DAS measurements. We employ the data obtained from the Sanriku seafloor observation system located near Iwate prefecture, in northeastern Japan (Fig. 1 ). This observatory is being operated by the Earthquake Research Institute, The University of Tokyo (Shinohara et al., 2019 ; 2022 ). The system consists of three three-component accelerometers and two tsunami-meters connected to a 120-km optic fiber cable. Two cables are installed in the observation system: the first one was installed in 1996 and performs data transmission and system control, and the second was installed in 2015. The cable installed in 1996 has a total of 12 optic fibers. Six of these are used for data transmission, and the rest of them are spare (dark) fibers, which are suitable for DAS measurements (Shinohara et al., 2022 ). In addition, because there are no signal repeaters through the system the DAS measurements can be performed to the end of the cable. Both cables are connected to a land station in Kamaishi, Japan (Shinohara et al., 2019 ; 2021; 2022 ). We propose a framework, called SVR-DAS, that performs earthquake detection and seismic phase picking. To detect earthquake signals, we apply the template matching method (Gibbons and Ringdal, 2006 ). Template matching measures the similarity between the target trace and a master signal (or template) using the correlation coefficient (CC). This method has been used for both inland (Li and Zhan, 2018 ) and seafloor DAS data (Miao et al., 2025 ). Here, we use a different implementation of template matching. Figure 2 a shows the workflow of our implementation. First, we use envelopes instead of waveforms for the search. The envelope shape is simpler than the waveform, which enhances the CC values. Second, we use an initial template to perform the search. The initial template is retrieved from the DAS data itself, and the detections from the first search are used as templates for subsequent searches. The search is performed using 20-s time windows with 50% overlap. Both template and DAS waveforms are band-pass filtered between 5 and 15 Hz to reduce ocean noise and suppress high frequency signals. Figure 2 b shows an example of the results from applying template matching with one template. The data corresponds to an earthquake from 2023-07-26 12:26:00 UTC (Fig. 2 b, middle panel). The template used is from channel 800 (Fig. 2 b, upper panel), located about 32 km from the landing station. This channel corresponds to the buried section of the cable. Along the cable, we observe a clear pattern of CC values above 0.8 (Fig. 2 c, middle and lower panels) that coincides with the time at which the earthquake signal appears in the DAS data. In the rest of this section, we present additional results obtained from the DAS data of this event. False detections may arise due to the similarity between the envelope and the later sections of DAS traces. However, we can verify whether the DAS data contains a signal by counting the number of false detections per file. The number of false detections caused by noise is relatively constant. Hence, the DAS data is considered to contain signals if we observe an increase in the number of detections above a mean value. Therefore, we set a detection threshold based on the mean number of false detections in DAS data. Through trial and error, we set this threshold to 30 counts. After detection, the algorithm proceeds to identify P and S arrival times. Figure 3 a shows the flow diagram that summarizes the seismic phase picking phase performed by SVR-DAS. First, we load two files: the file that contains all detections obtained by template matching, and the original file. The data is once again band-pass filtered between 5–15 Hz due to the algorithm stores unfiltered detections. In addition, we increase the time length of the detections from 20 to 30 s and estimate the SNR of all detections. We assume that the P wave may be missed within the 20-s time window if the SNR is relatively low. Hence, we change to 30-s time window length. Otherwise, P phase may be within the time windows, and the algorithm use the 20-s time windows instead. SVR-DAS applies two picking methods to identify P and S in the DAS data: recursive STA/LTA (Allen, 1982 ), and auto regressive models using the Akaike Information Criterion (AR - AIC) (Akazawa, 2004 ). Each method has its own fixed parameters, and these are summarized in Table S1 from Supplemental Material. Figure 3 b shows two DAS traces obtained from 2023-07-26 data and their corresponding P and S picks obtained using STA/LTA and AR-AIC. Good agreement can be observed in P and S picks obtained from the two methods. Pick disagreements between methods may arise due to the variable SNR between traces. Hence, the picks obtained from these methods are compared with each other, and a pick coherency threshold is set to discriminate between coherent and non-coherent picking. By coherent picking, we refer to the agreement between pick estimations. We set the pick threshold to 3 s, and the initial pick is set as the mean both picks if the time difference between picks is below the threshold. On the contrary, if the separation between peaks is above 3 s, then the algorithm discards the picks. The pick coherency threshold should filter most of the incoherent picks. However, it is still plausible the case where all methods agreed in a wrong pick estimation. For this reason, after obtaining initial picks, the algorithm performs an outlier removal process that consists of the evaluation of the slope changes between adjacent picks obtained from the DAS traces. We take advantage of the spatial information of DAS data, assuming that coherent picking between channels must have minimum slope changes. If no outliers are present, then the slope change between picks is constant through all points. In contrast, the outliers may change abruptly the slope. The algorithm estimates the slope for all obtained picks and discards the pickings that separate from the mean slope. After applying this procedure 2023-07-26 DAS data, we observe that most of the obtained P and S picks align agrees with the apparent arrival times observed in the DAS data (Fig. 4 a). Up to this point we may obtain reasonable estimations of P and S arrival times. To further improve the estimation of P and S picks we apply Support Vector Regression (SVR), a regression method based on Support Vector Machines (SVM). SVM is a supervised Machine Learning (ML) method used for classification, regression and outlier detection, and consists of the search of hyperplanes in higher dimensional spaces that best separates the data (Vapnik, 1997 ; Pedregosa et al., 2012). The application of SVM in regression problems is the SVR method, which searches best hyperplanes that fits the data. Linear and nonlinear regression is possible due to kernel search instead of model parameters search (Smola and Scholkopf, 2003; Pedregosa et al., 2012). An advantage of SVR is that only a subset of training dataset is needed to perform the regression. For more details about the theoretical aspects of SVM and SVR, the reader can refer to Smola and Scholkopf (2003). We implement SVR algorithm provided by the Python package Scikit-Learn (Pedregosa et al., 2012). The kernel selection is based on two parameters: the density of P and S picks through the data, and the interquartile range (IQR) of P and S picks. By trial and error, we use linear regression to the data with IQR above than 2.1. In addition, the linear regression is also set to the data where the density of picks is concentrated in a small section of DAS data. Otherwise, we set Radial Basis Functions (RBF) as the kernel to perform the regression. The reason for this choice is that data with small IQR may be associated with events that occurred near the cable. Thus, the P and S picks behavior may present a nonlinear trend. P-to-S conversions are observed in ocean-bottom seismometers and DAS measurements due to the sediment layers in the Sanriku area (Fujie et al., 2018 ; Fukushima et al., 2022 ). Because DAS is prone to record strain in horizontal direction due to the aspect ratio of the optic fiber, it is most likely that the amplitude of the P-to-S converted wave is more prominent than P waves. Assuming that the model for P may be related P-to-S converted waves, we need to refine our picking to capture as accurately as possible the P wave arrival time. Hence, we adopt a procedure based on the signal spectrogram to refine the P picks obtained from the SVR model. The spectrogram is calculated in a 20-s time window centered in the estimated P pick. If there are no other impulsive signal or noise inside the time windows, it is probable that we observe in the spectrogram a change in the power of the signal due to the P wave. Then, we create envelope functions using the obtained spectrograms. The envelope functions are normalized with respect to the maximum value of the spectrogram. We also attempt to refine S wave picking using this approach if there is a moveout due to the SVR regression. An amplitude threshold is set to retrieve both refined P and S arrival times. For P wave, the threshold is set as the 20% of the median absolute deviation (MAD) of the envelope function. Because the P wave amplitude decreases at the relatively far section of the cable, the threshold changes to 10% of the MAD value. For S wave, the threshold is set as the 80% of the MAD of the envelope. An example of this envelope function, and the pickings obtained using this procedure is shown in Figure S1 from Supplemental Material. Most of the picking is automatic and self-guided by the algorithm. Nevertheless, a manual revision is made for all pickings. At this point, P and/or S pickings may be discarded or readjusted. The final catalog is stored in an HDF5 file following the STEAD database format proposed in Mousavi et al. ( 2020 ). We evaluate the performance of the method by comparing the SVR-DAS pickings with manual picks obtained of 400 channels from DAS data from 2023-07-26. These channels lie within the first 80-km DAS cable section (see Fig. 1 ). The evaluation of the picks is summarized in Table 1 . Figure 5 shows the histogram and the comparison between SVR-DAS and manual picks. We compare both datasets by estimating the standard deviation (σ) and the IQR of P and S picks. The manual choice of P and S arrival times across the channels must be lower than the obtained from picking algorithms. The difference between manual picking and SVR-DAS of IQR and σ is around 0.1 and 0.6 except for the IQR of P picks, which increases to 1.92. The difference of arrival times between manual and SVR-DAS for most of P and S picks is within ± 1 s. For P, we observe that the picks distribution roughly coincides up to 40 km. Above this point, the pick distributions associated with manual picks are more widespread. This is reasonable due to the decrease in SNR for the farthest channels. In contrast, the distributions of S picks do not vary significantly for most of the data. It seems that there is less uncertainty in the choice of S wave arrival times. However, manual picking was hard for most of the channels that correspond to the cable section above 40 km. Table 1 Comparison of the standard deviation (σ) and interquartile range (IQR) between manual and SVR-DAS picks. Value σ IQR t SVR P 2.63 5.02 t M P 2.22 3.77 t SVR S 3.89 3.69 t M S 6.69 6.59 (t S - t p ) SVR 1.36 1.94 (t S - t p ) M 1.66 2.43 Results We applied SVR-DAS to the data from 2022-02-28 08:51:20 to 2022-02-28 23:59:20 h. All data is in UTC format. Figure 6 shows the results obtained from template matching algorithm. The data was acquired using a gauge length of 100 m, and the spatial sampling is 16.44 m. The acquisition sampling rate was set to 800 Hz; however, we decimated the DAS data to 100 Hz. We set as the initial template the same DAS trace from 2023-07-26 12:26:02 h (see Fig. 2 ). The algorithm identified 23 events with mean SNR values ranging from 0.03 dB to 0.27 dB (Fig. 6 a). The CC values of all detections range between 0.8 and 0.9. The number of detections varies between 3 and 870, with a mean value of 25 detections per DAS file. We plotted the DAS data and their associated CC values of three identified events (Fig. 6 b). For plotting, we increased the window size of the DAS data to 180 s to avoid incomplete events. The blue dotted line in each panel of DAS data highlights the original file size used in the template matching search. We observe that the highest values of CC coefficients coincide with the sections of DAS data where the highest amplitudes are present. The values that coincide with the DAS data at 60 s are discarded when retrieving the detections from the data. Not all detections are associated with earthquakes. We can identify the unrelated signals and separate them from the earthquake signals using the standard deviation of the CCs. From the 23 detections obtained from template matching, 8 of these detections contained earthquake-related strain rate signals with arrival time differences S – P above 20 s, and 3 are pure noise. From the remaining events, the algorithm estimated P and S picks from 8 events, and the other 4 most of the pickings were from S waves. We build the training and validation dataset using the P and S pickings from these 8 events. A comparison between initial picks obtained from STA/LTA and AR-AIC methods is shown in Fig. 7 . We show six DAS traces from 2022-02-28 10:16:20 UTC (Fig. 7 a), 2022-02-28 10:31:20 UTC (Fig. 7 b) and 2022-02-28 12:01:20 UTC (Fig. 7 c). In the left column, we show the traces where there is well agreement between the picks, and in the right column we show the traces with picks that do not fulfill the pick coherency threshold criterion. For relatively high SNR signals both STA/LTA have good agreement. AR-AIC method is more robust to the presence of noise than STA/LTA. However, mispicking may arise in data with non-earthquake impulsive signals (Küperkoch et al., 2011 ). For non-coherent picks, we observe that the difference between obtained picks for P, S picks or both ranges from 4 to 15 s. This range may increase in other channels if the SNR is relatively lower (see Figures S2-S3 from Supplemental Material). The training points for SVR and the obtained pick models for P and S are shown in Fig. 8 . The number of training examples is not fixed and varies with the available number of data points. In addition, the parameter for obtaining the model differs depending on the distribution of points and the SNR. For the highest SNR examples, we use an RBF kernel with regularization parameter C of 1.0 and an epsilon-tube parameter ε of 0.1. After several trial-and-error tests, we set these parameters which are the default values for this method (Pedregosa et al. 2011). However, if we use a linear kernel we decrease both C and ε parameters to avoid noncoherent solutions. Thus, we set C to 0.1 and 0.05 for linear fitting for P and S models, respectively. The pick models obtained from SVR are consistent with the observed P and S pattern in the DAS data. Figure S4 from Supplemental Material shows another example of a detected event with their associated P and S picking models. Although, we can observe a good fit between the results and the observed data some points must be considered when using SVR. First, the models may be unstable with low-SNR data. We observe in the P and S models from 2022-02-28 12:01:20 UTC a clear deviation of the P model starting after 40 km. This observation coincides with an increase in noise in the DAS data. A better fit may be achieved if we decrease the C parameter from the SVR model. However, it may force the model to fit a line, producing a non-realistic fitting. Therefore, we could see more model misfits at the channels with lower SNR. The obtained results from SVR-DAS are then used as training and validation datasets to train a deep learning model. For this purpose, we use a CNN-LSTM-based model (Mendo-Pérez et al., 2025 ) for sequential data. The architecture of this model is based on the ones used in trigger voice systems (Supriya et al., 2020 ). As the model input, we use the spectrogram of the individual DAS traces, calculated using 0.6-s time windows with an 95% overlap. Prior to create the spectrograms, the data is downsampled from 100 Hz to 33.3 Hz. In addition, the data is normalized by subtracting the mean and dividing it to the standard deviation. To train this model, we use a subset of 18000 DAS data labels from our catalog. From these, 12000 examples are associated with earthquake signals, and 6000 examples are noise. The subset is separated into training and validation datasets using a ratio of 8:2. We use the Mean Squared Loss, as the loss function, and for backpropagation stage we use the AdamW optimizer using a fixed learning rate of 0.001. We apply a 4-s width Gaussian masks to create the output labels. The maximum probability of the masks is set to 1.0 at the time position of each P and S arrival time obtained from SVRDAS. The model is trained using 20 epochs and a batch size of 8. The model took around 1 hour to complete training using an NVIDIA GeForce RTX 3060. Figure 9 shows both the loss and accuracy curves. Here, we define accuracy of the model as the ratio of the number of labels correctly obtained by the model and the total number of labels. Assuming the accuracy is 100% if the model correctly identifies all examples, the accuracy of the model at the end of the training stage was 96% for both training and validation datasets. The obtained labels from the CNN-LSTM model can be seen in Fig. 10 . Here, we show four random DAS traces and their associated spectrograms retrieved from the validation datasets. We compare the labels obtained from the SVR-DAS method and the labels obtained by the CNN-RNN model assuming that the SVR-DAS labels are our ground truth. The maximum peaks of both P and S labels coincide at the same time as the ground truth labels. The probabilities for P and S label are 0.85–0.94 and 0.96–0.97, respectively. We observe that the trained model is prone to identify better the S arrival times than P. This problem may be resolved by adding more training examples of well-defined P arrival times in DAS traces with variable SNR. We also compare the arrival times obtained from SVR-DAS to the ones obtained by PhaseNet-DAS (Zhu et al., 2023 ), a deep-learning-based method for seismic phase picking, to estimate P and S arrival times. This model follows a semi-supervised scheme by using a pretrained PhaseNet model (Zhu and Beroza, 2019 ), to estimate initial DAS picks, followed by phase association using GaMMA (Zhu et al., 2022 ) to filter non-coherent picks. The obtained picks are used to train a UNet-based model similar to PhaseNet (Rommberg et al., 2015) to estimate P and S arrival times in DAS data. Figure 11 shows the picks obtained from three events: 2022-02-28 10:16:20 h (Fig. 11 a), 2022-02-28 12:01:20 h (Fig. 11 b), and 2023-07-26 12:16:03 h (Fig. 11 c). From this point, we will refer to the events as E1, E2, and E3 for simplicity. In these three events, both methods capture most of the arrival times with apparently the same distribution but present some noteworthy differences. Note that PhaseNet – DAS in E1 also detects a small event between 20–40 s before to the main event (Fig. 11 a). The reason that SVR-DAS did not identify this event is due to the choice of the time windows to run and STA/LTA and AR-AIC pickers. SVR-DAS is thought to identify one event per time window. Hence, a disadvantage in SVR-DAS is that it cannot identify more than one earthquake per time window. To detect consecutive events, smaller time window must be used to fulfill the one event condition. PhaseNet – DAS estimated P picks up to 80 km coinciding with a decrease in SNR. Using SVR-DAS we can obtain P picks in the low SNR region due to the SVR fitting. In E2 we observe that the P pick estimation stops near 100 km, where the noise starts to increase. In this case, SVR-DAS also provide smoother results. The S pick distributions of both methods in both events coincide very well. Finally, in E3, we observe that there is a small gap in PhaseNet-DAS P picks between 60 and 80 km. Another feature is the distribution of P picks around 20 s with a slope nearly horizontal. A closer inspection to each one of the DAS traces was done shows that this feature seems related to a non-physical event. In contrast, SVR-DAS P picks show a continuous distribution with no gaps. To quantitatively evaluate the picking performance of SVR-DAS, we estimate the time arrival difference obtained between SVR-DAS and both PhaseNet-DAS (t SVR - t PN ) and manual picking (t SVR – t M ) for P and S phases. The histograms associated with these quantities are shown in Fig. 12 . In addition, Table 2 shows the mean ( µ ), median, minimum (min) and maximum (max) values, standard deviation ( σ ), interquartile range (IQR), and Pearson correlation coefficient (R) for the time differences of P and S picks. For simplicity, we use the notation Δt PN P, Δt M P, Δt PN S and Δt M S to refer to the difference between SVR-DAS and PhaseNet-DAS for P phase, SVR-DAS and manual picking for P phase, SVR-DAS and PhaseNet-DAS for S phases, and SVR-DAS and manual picking for S phase, respectively. The distributions for P and S phases associated with manual picking are more widespread than the ones associated with PhaseNet-DAS. According to Table 2 , the difference between the standard deviations (Δσ) of P picks distributions associated with PhaseNet-DAS and manual picking for E1, E2 and E3 are 1.08,1.97, and 0.18, respectively. The IQR difference (ΔIQR) for is 0.75 for E1, 0.45 for E2, and 0.21 for E3. In the case of S picks, Δσ is 0.38 for E1, 1.35 for E2, and 0.26 for E3. For ΔIQR, the values of E1, E2 and E3 are 0.14, 0.98, and 0.34. We observe that σ and IQR values for the three events are higher for E1 and E2. Although for E3 is practically the same tendency, the difference is smaller than the observed in the other events. The central values of the distributions in E1 and E3 are prone to the left, meaning that PhaseNet-DAS picks and manual arrival times estimations are after SVR-DAS arrival times. In the case of E2, we see that the central value of the P distribution is near 0 s, and the value of S distribution is above 0 s. Another observation is related to the correlation coefficient R between picks. For E1 and E3, we observe R values of PhaseNet-DAS and manual picking is equal or higher than 0.90 for both P and S picks. In contrast, the R values of E2 varies depending on the method. While for PhaseNet-DAS-related picks R is 0.77 s and 0.32 s, for manual picking the values are − 0.16 and 0.10. With all these elements, we can conclude that picking of SVR-DAS and PhaseNet-DAS is consistent in these events. In contrast, manual picking presents wider distributions due to the bias in the time arrival picking of P and S in relatively lower SNR DAS channels. Particularly, manual picking in E2 was challenging. Although in the DAS data you can clearly see the P and S time arrivals, it is difficult to select accurately in each channel due to the low SNR. Table 2 Comparison of mean (µ), median, minimum value (Min), maximum value (Max), standard deviation (σ), and interquartile range (IQR) estimated from the difference in arrival times of P (Δtp) and S (Δtp) picks obtained from SVR-DAS, PhaseNet-DAS, and manual picking. 2022-02-28 10:16:20 h (E1) Parameter µ median min max σ IQR R Δt PN P -0.52 s -0.54 s -0.92 s 0 s 0.16 s 0.15 s 0.99 s Δt M P -1.56 s -1.39 s -6.65 s 1.78 s 1.24 s 0.90 s 0.90 s Δt PN S -0.46 s -0.43 s -1.17 s 0.10 s 0.30 s 0.44 s 0.98 s Δt M S -0.53 s -0.48 s -3.70 s 4.33 s 0.68 s 0.58 s 0.99 s 2022-02-28 12:01:20 h (E2) Δt PN P -0.12 s -0.08 s -0.75 s 0.28 s 0.29 s 0.51 s 0.77 s Δt M P -0.33 s 0.11 s -11.59 s 3.26 s 2.26 s 0.96 s -0.16 s Δt PN S 0.45 s 0.45 s -0.47 s 1.51 s 0.51 s 0.79 s 0.39 s Δt M S 1.16 s 1.43 s -13.72 s 4.37 s 1.86 s 1.77 s 0.10 s 2023-07-26 12:26:03 h (E3) Δt PN P -0.84 s -0.57 s -3.20 s 2.16 s 0.88 s 0.76 s 0.98 s Δt M P 0.40 s 0.20 s -2.63 s 2.66 s 0.70 s 0.97 s 0.96 s Δt PN S -0.26 s -0.31 s -2.25 s 0.46 s 0.31 s 0.31 s 0.99 s Δt M S 0.28 s 0.15 s -2.14 s 2.71 s 0.57 s 0.65 s 0.98 s Discussion Here, we present a ML-based framework to obtain training and validation datasets from seafloor DAS measurements. We demonstrate that this method is useful to identify earthquake signals using a simple template, and to identify P and S arrival times from scratch using common seismic picking methods. The novelty of this method relies in the recovery of P and S pickings from most of the DAS channels using SVR. The uncertainty of the pickings is based on the SNR of each channel. The availability of DAS-based databases is vital for developing deep learning models for seafloor DAS data. Self-noise in DAS data is due to signal fading, laser noise, and common-mode noise (Lindsey et al., 2020 ; Farghal et al., 2022 ). Additionally, seafloor DAS measurements present additional sources of noise. Common 2D filtering techniques are useful to identify and extract earthquake signals from strain rate measurements. This has been done to identify surface waves to obtain P and S velocity models (Spica et al., 2020 ; Fukushima et al., 2022 ; 2025 ). However, nonphysical signals may have lie in the same frequency band as the extracted signal. The source of this noise may lie in the cable coupling into the seafloor. Most of the measurements are done using telecommunication fibers deployed for signal transmission. Hence, the coupling may not be the same throughout the cable (Farghal et al., 2022 ). It has been demonstrated that the strain amplitude can change significantly with the degree of coupling of the cable (Ajo-Frankin et al., 2019; Lindsey and Martin, 2020 ). For this reason, DAS-based datasets for deep learning must address these types of noise. Although our method has been proven useful, there are some limitations. First, the kernel choice for SVR is based on the number of training points distributed throughout the DAS channels, and the interquartile range of the difference between the arrival times of P and S. Although the examples shown demonstrated that this criterion works, there may be events that do not fulfill these criteria. Current work is based on analyzing more DAS data to establish a more robust criterion for the kernel selection. Deeping more in the kernel’s discussion, so far, we have Radial Basis Functions (RBF) and linear kernels for obtaining picking models. The main reason for using these two are based on the following points: 1) Time calculations, and 2) distribution of pickings across the DAS data. Earthquake events whose source is close to the cable will produce nonlinear arrival times distribution, whereas farther events will display a linear distribution with different slopes due to the different apparent velocities of P and S. The time execution of SVR algorithms depends on the number of points, but it ranges between 0.3–5.0 s. Linear kernels are more time expensive than RBF kernels. It is possible that other kernels such as polynomial, sigmoid, etc. may produce a better adjustment. However, this may result in an increase in the time calculation. Another limitation lies in the picking methods. Overall, STA/LTA and AR-AIC methods have good agreement in relatively high SNR signals and disagrees in when decreasing the SNR. The reason for the agreement is that prior to the estimation of the autoregressive models, the method applies STA/LTA for P and S phrases (Akazawa et al., 2004). However, we found events such as 2022-02-28 16:52:20 UTC that failed to predict. Although we observe in the DAS section (see Figure S5 – S6 from Supplemental Material) two clear arrivals, the pickings obtained from both methods completely mismatch. Further research will be done to obtain a more robust implementation of this framework. Conclusions Here we present SVR-DAS, a ML-based method to create training and validation datasets from seafloor Distributed Acoustic Sensing measurements. Our method is divided into two stages: earthquake detection and seismic phase picking. In the first stage, our method applies envelope template matching to identify earthquake strain rate measurements in DAS channels. The template matching can start from one template, and the algorithm feedback the template library with the new detections. In the second stage, we apply STA/LTA (Allen, 1982 ) and autoregressive models using the Akaike Information Criterion (AR-AIC) (Akazawa et al., 2014) to pick P and S arrival times from the DAS channels. These pickings are compared to each other, and the coherent pickings are used as training sets to obtain P and S pick models using Support Vector Regression (SVR) (Vapnik, 1997 ). Because P-to-S converted waves have higher amplitude than the true P wave, most of the pickings are associated with the converted wave. For this reason, we apply a spectrogram-based picker using a time window near the P-to-S converted waves in order to identify the time where the signal power increases due to the P wave arrival. We tested our method using relatively high SNR DAS data associated with an earthquake that occurred in 2023-07-23 successfully retrieving both P and S arrival times. Afterwards, we applied SVR-DAS using the data from 2022-02-28 08:51:20 h to 23:59:20 h UTC. Our method identified 9 events with difference in P and S time arrivals less than 20 s. From these events, we retrieved approximately around 450 000 DAS channels with identified P and S time arrivals. Using a subset of 18 000 examples (12 000 events and 6000 noise), we train a CNN-RNN model to identify P and S seismic pickings, achieving an accuracy of 96%. We believe that the generated DAS-based databases obtained from SVR-DAS will be useful to develop state-of-the-art deep learning models, or fine tuning already existent deep learning models, to explore the seafloor DAS measurements. A potential application of this method is also in Earthquake Early Warning system, that relies on accurate P wave estimation to determine ground motion shaking. Declarations Acknowledgments This study is supported by the MEXT project for Seismology Toward Research Innovation with Data of Earthquakes (STAR-E) No. JPJ010217. The key ideas in this study were derived from the activities of the MEXT Volcano Practical Human Resource Development Support Program Japan, the MEXT The Third Earthquake and Volcano Hazards Observation and Research Program (Earthquake and Volcano Hazard Reduction Research), JST NEXUS, Japan No. JPMJNX25C5, the JSPS KAKENHI Grant-in-Aid for Scientific Research (A) No. 23H00466, Fund for the Promotion of Joint International Research (International Collaborative Research) No. 23KK0181, Grant-in-Aid for Challenging Research (Pioneering) No. 25K21806, and Earthquake Research Institute, The University of Tokyo Joint Research ERI JURP 2025-A-03, 2024-B-01, and 2025-B-01. We thank Dr. Shinichi Ito, Dr. Tomoki Tokuda, and Hiroaki Yamahana for their valuable feedback regarding this study. We thank T. Yagi from ERI, The University of Tokyo for lending us the data on Sanriku cable. References Allen R (1982) Automatic phase pickers: Their present use and future prospects. Bull. Seis. Soc. Am. 72(6B): S225 – S242 Ajo-Franklin JB, Dou S, Lindsey NJ, Monga I, Tracey C, Robertson M (2019) Distributed acoustic sensing using dark fiber for near surface characterization and broadband seismic event detection. Sci Rep 9(1):1328. 10.1038/s41598-018-36675-8 Akazawa T (2004) A technique for automatic detection of onset time of P- and S- Phases in strong motion records Aoi S, Asano Y, Kunugi T, Kimura T, Uehira K, Takahashi N, Ueda H, Shiomi K, Matsumoto T, Fujiwara T (2020) MOWLAS: NIED observation network for earthquake, tsunami and volcano. Earth Planet Space 72(126). 10.1186/s40623-020-01250-x Aoi S, Takeda T, Kunugi T, Shinohara M, Miyoshi T, Uehira K (2023) Development and Construction of Nankai Trough Seafloor Observation Network for Earthquakes and Tsunamis: N-Net. 2023 IEEE Underwater Techonology. 10.1109/UT49739.2023.10103206 Atterholt J, Zhan Z, Shen Z, Li Z (2022) A unified wavefield-partitioning approach for distributed acoustic sensing. Geophys J Int 228(2):1410–1418. 10.1093/gji/ggab-407 Baba S, Araki E, Yokobiki T, Kawamata K, Uchiyama K, Yoshizuka T (2024) Seismic observation using distributed acoustic sensing around the Tsugaru Strait at the Japan and Kuril Trenches, northeastern Japan. Earth Planet Space 76(28). 10.1186/s40623-024-01975-z Benioff H (1935) A linear strain seismograph. Bull Seism Soc Am 25(4):283–309 Castongia E, Wang HF, Lord N, Fratta D, Mondanos M, Chalari A (2017) An experimental investigation of distributed acoustic sensing (DAS) on lake ice. J Environ Eng Geophys 22:167–176. 10.2113/JEEG22.2.167 Daley TM, Freifeld BM, Ajo-Franklin J, Dou S, Pevzner R (2013) Field testing of fiber-optic distributed acoustic sensing (DAS) for subsurface seismic monitoring. Lead Edge 32(6):593–724. 10.1190/tle32060699.1 Ding Q, Shen Z, Zhu W, Liu B (2025) DASFormer: self-supervised pretraining for earthquake monitoring. Visual Intell 3(14). https://doi.org/10.1007/s44267-025-00085-y Farghal NS, Saunders JK, Parker GA (2022) The Potential of Using Fiber Optic Distributed Acoustic Sensing (DAS) in Earthquake Early Warning Applications. Bull Seism Soc Am 115(2):516–532. 10.1785/0120240234 Fujie G, Kodaira S, Kaiho Y, Yamamoto Y, Takahashi T, Miura S, Yamada T (2018) Controlling Factor of incoming plate hydration at the north-western Pacific margin. Nat Commun 9(3844). 10.1038/s41467-018-06320-z Fukushima S, Shinohara M, Nishida K, Takeo A, Yamada T, Yomogida K (2022) Detailed S-wave velocity structure of sediment and crust off Sanriku, Japan by a new analysis method for distributed acoustic sensing data using a seafloor cable and seismic interferometry. Earth Planet Space 74(92). 10.1186/s40623-022-01652-z Fukushima S, Shinohara M, Yamada T, Hino R, Azuma R, Ito Y, Yamashita Y, Takano H (2025) Enhanced P-wave velocity imaging by marine airgun-source seismic surveys with distributed acoustic sensing. Sci Rep 15(18111). 10.1038/s41598-025-01190-0 Gibbons SJ, Ringdal F (2006) The detection of low magnitude seismic events using array-based waveform correlation. Geophys J Int 165(1):149–166. 10.1111/j.1365-246X.2006.02865.x Hartog AH (2017) An Introduction to Distributed Optical Fibre Sensors. CRC, Boca Raton Hough SE (1990) Constraining sediment thickness in the San Francisco Bay Area Using Observed Resonances and P-to-S Conversions. Geophys Res Lett 17(9):1469–1472. 10.1029/GL017i009p01469 Kaneko R, Nagao H, Ito S, Tsuruoka H, Obara K (2023) Detection of Deep Low-Frequency Tremors From Continuous Paper Records at a Station in Southwest Japan About 50 Years Ago Based on Convolutional Neural Network. J Geophys Res Sol Earth 128(2). 10.1029/2022JB024842 Kanazawa T, Uehira K, Mochizuki M, Shinbo T, Fujimoto H, Noguchi S, Kunugi T, Shiomi K, Aoi S, Matsumoto T, Sekiguchi S, Okada T (2016) S-net project, cabled observation network for earthquakes and tsunamis. SubOptics 2016, WE2B-3 Katoh S, Iio Y, Nagao H, Katao H, Sawada M, Tomikasa K (2025) SegPhase: development of arrival time picking models for Japan’s seismic network using the hierarchical vision transformer. Earth Planet Space 77(118). 10.1186/s40623-025-02249-y Katoh S, Nagao H, Iio Y (2025) PoViT-UQ: P-wave Polarity and Arrival Time Determination using Vision Transformer with Uncertainty Quantification. Geophys J Int. 10.1093/gji/ggaf324 Kawaguchi K, Araki E, Kaneda Y (2011) Establishment for a method for real-time and long-term seafloor monitoring. J Adv Mar Sci Technol Soc 17:125–135. 10.14928/amstec.17.2_125 Küperkoch L, Meier T, Diehl T (2011) Automated Event and Phase Identification. In: Bormann, P. (Ed.), New Manual of Seismological Observatory Practice 2 (NMSOP-2), Postdam: Deutsches GeoForschungsZentrum GFZ, 1–52. 10.2312/GFZ.NMSOP-2_ch16 Li Z, Zhan Z (2018) Pushing the limit of earthquake detection with distributed acoustic sensing and template matching: a case study at the Brady geothermal field. Geophys J Int 215(3):1583–1593. 10.1093/gji/ggy359 Lindsey NJ, Martin ER, Dreger DS, Freifeld B, Cole S, James SR, Biondi BL, Ajo-Franklin JB (2017) Fiber-Optic Network Observations of Earthquake Wavefields. Geophys Res Lett 44(23):11792–11799. 10.1002/2017GL075722 Lindsey NJ, Martin RM (2020) Fiber-optic seismology. Annu Rev Earth Planet Sci 49:309–336. 10.1146/annurev-earth-072420-065213 Lindsey NJ, Rademacher H, Ajo F (2020) On the broadband instrument response of fiber-optic DAS arrays. J. Geophys. Res. 125(e2019JB018145). 10.1029/2019JB018145 Liu B, Yang S, Ren Y, Xu X, Jiang P, Chen Y (2021) Deep-learning seismic full-waveform inversion for realistic structural models. Geophysics 86(1). https://doi.org/10.1190/geo2019-0435.1 Matsubara M, Obara K, Kasahara K (2008) Three-dimensional P- and S-wave velocity structures beneath the Japan Islands obtained by high-density seismic stations by seismic tomography. Tectonophysics 454(1–4):86–103. 10.1016/j.tecto.2008.04.016 Masoudi A, Newson TP (2016) Contributed review: distributed optical fibre dynamic strain sensing. Rev Sci Instrum 87:011501. 10.1063/1.4939482 McBrearty I, Beroza GC (2023) Earthquake Phase Association with Graph Neural Networks. Bull Seism Soc Am 113(2):524–547. 10.1785/0120220182 Miao Y, Salaree A, Spica ZJ, Nishida K, Yamada T, Shinohara M (2025) Assessing the Earthquake Recording Capability of an Ocean-Bottom Distributed Acoustic Sensing Array in the Sanriku region, Japan. Seis Res Lett 96:631–650. 10.1785/0220240120 Mendo-Pérez G, Nagao H, Katoh S, Shinohara M (2025) Leveraging CNN-RNN-based model for earthquake detection in Distributed Acoustic Sensing measurements [Work in preparation]. The University of Tokyo Mousavi SM, Ellsworth WL, Zhu W, Chuang LY, Beroza GC (2020) Earthquake transformer – an attentive deep-learning model for simultaneous earthquake detection and phase picking. Nat Commun 11(3952). 10.1038/s41467-020-17591-w Moseley B, Nissen-Meyer T, Markham A (2020) Deep learning for fast simulation of seismic waves in complex media. Solid Earth 11(4):1527–1549. https://doi.org/10.5194/se-11-1527-2020 Muir JB, Zhan Z (2022) Wavefield-based evaluation of DAS instrument response and array design. Geophys J Int 229(1):21–34. 10.1093/gji/ggab439 Ross ZE, Meier MA, Hauksson E, Heaton TH (2018) Generalized Seismic Phase Detection with Deep Learning. Bull Seismo Soc Am 108(5A):2894–2901. https://doi.org/10.1785/0120180080 Ross ZE, Yue Y, Meier MA, Hauksson E, Heaton TH (2019) PhaseLink: A Deep Learning Approach to Seismic Phase Association. J Geophys Res Sol Ea 124(1):856–869. https://doi.org/10.1029/2018JB016674 Shinohara M, Yamada T, Akuhara T, Mochizuki K, Sakai S, Hamakawa M, Kasajima T, Arioka T, Kubota S (2019) Distributed Acoustic Sensing measurement by using seafloor optical fiber cable system off Sanriku for seismic observation. Paper presented at OCEANS 2019 MTS/IEEE SEATTLE. 10.23919/OCEANS40490.2019.8962757 Shinohara M, Yamada T, Akuhara T, Mochizuki K, Sakai S (2022) Performance of a Seismic Observation by Distributed Acoustic Sensing Technology Using a Seafloor Cable Off Sanriku, Japan. Front Mar Sci 9. 10.3389/fmars.2022.844506 Spica ZJ, Nishida K, Akuhara T, Petrelis F, Shinohara M, Yamada T (2020) Marine Sediment Characterized by Ocean-Bottom Fiber-Optic Seismology. Geophys Res Lett 47(16). 10.1029/2020GL088360 Spica ZJ, Castellanos JC, Viens L, Nishida K, Akuhara T, Shinohara M, Yamada T (2022) Subsurface Imaging With Ocean-Bottom Distributed Acoustic Sensing and Water Phases Reverberations. Geophys Res Lett 49(2). 10.1029/2021GL095287 Supriya K, Divya A, Vinodkumar B, Ram Sai G (2020) Trigger Word Recognition using LSTM. Int J Eng Res Technol 9(6) Takahashi N, Kodaira S, Tsuru T, Park J, Kaneda Y, Suyehiro K, Kinoshita H, Abe S, Nishino M, Hino R (2004) Seismic structure and seismogenesis off Sanriku region, northeastern Japan. Geophys J Int 159(1):129–145. 10.1111/j.1365-246X.2004.02350.x Tokuda T, Nagao H (2023) Seismic-phase detection using multiple deep learning models for global and local representations of waveforms. Geophys J Int 235(2):1163–1182. 10.1093/gji/ggad270 Vapnik VN (1997) The support vector method. In International conference on artificial neural networks (pp. 261–271). Berlin, Heidelberg. Springer Berlin Heidelberg Viens L, Bonilla LF, Spica ZJ, Nishida K, Yamada T, Shinohara M (2022) Nonlinear Earthquake Response of Marine Sediments With Distributed Acoustic Sensing. Geophys Res Lett 49(21). 10.1029/2022GL100122 Walter F, Graff D, Lindner F, Paitz P, Koppfli M, Chmiel M, Fichtner A (2020) Distributed acoustic sensing of microseismic sources and wave propagation in glaciated terrain. Nat Commun 11:2436. 10.1038/s41467-020-15824-6 Xiao H, Tilmann F (2024) DeepDAS: An Earthquake Phase Identification Tool Using Submarine Distributed Acoustic Sensing Data, Galileo conference: Fibre Optic Sensing in Geosciences, Catania, Italy, 16–20 Jun 2024, GC12-FibreOptic-51. https://doi.org/10.5194/egusphere-gc12-fibreoptic-51 Yin J, Zhu W, Li J, Biondi E, Miao Y, Spica ZJ, Viens L, Shinohara M, Ide S, Mochizuki K, Husker AL, Zhan Z (2023) Earthquake Magnitude With DAS: A Transferable Data-Based Scaling Relation. Geo Phys Lett 50(10). 10.1029/2023GL103045 Yu S, Ma J, Wang W (2019) Deep learning for denoising. Geophysics 84(6). https://doi.org/10.1190/geo2018-0668.1 Yu C, Zhan Z, Lindsey NJ, Ajo-Franklin JB, Robertson M (2019) The Potential of DAS in Teleseismic Studies: Insights From the Goldstone Experiment. Geophys Res Lett 46(3):1320–1328. 10.1029/2018GL081195 Zhan Z (2019) Distributed Acoustic Sensing. Turns Fiber-Optic Cables into Sensitive Seismic Antennas. Seism Res Lett 91(1):1–15. 10.1785/0220190112 Zhu W, Beroza GC (2019) PhaseNet: a deep-neural-network-based seismic arrival-time picking method. Geophys J Int 216(1):261–273. 10.1093/gji/ggy42 Zhu W, Biondi E, Li J, Ross ZE, Zhan Z (2023) Seismic arrival-time picking on distributed acoustic sensing data using semi-supervised learning. Nat Commun 14(8192). 10.1038/s41467-023-4335-3 Zhu W, McBrearty I, Mousavi SM, Ellsworth WL, Beroza GC (2022) Earthquake Phase Association Using a Bayesian Gaussian Mixture Model. J Geophys Res Sol Earth 127(5). 10.1029/2021JB023249 Zhang J, Li Z, Zhang J (2023) Simultaneous Seismic Phase Picking and Polarity Determination with an Attention-Based Neural Network. Seism Res Lett 94(2A):813–828. 10.1785/0220 Additional Declarations The authors declare no competing interests. Supplementary Files MendoPerezetalDASSuppFile.docx Supplementary Information Cite Share Download PDF Status: Posted Version 2 posted You are reading this latest preprint version Show more versions Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8080988","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":543657652,"identity":"af0e1adf-5c67-4447-b9cc-2f483ddea9c4","order_by":0,"name":"Gerardo Mendo-Pérez","email":"","orcid":"https://orcid.org/0000-0002-5239-4377","institution":"Earthquake Research Institute, The University of Tokyo","correspondingAuthor":false,"prefix":"","firstName":"Gerardo","middleName":"","lastName":"Mendo-Pérez","suffix":""},{"id":543657653,"identity":"753654f7-fa71-4c3b-95f1-b81463b43e71","order_by":1,"name":"Hiromichi Nagao","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABBUlEQVRIiWNgGAWjYLACxgYbBgZmCNsAKsZGSEsaVEsC8VoOQ1kILbiBOfvxix8+7jgvZ97OwCbB+MPG2OB4A+OHHwx8ebi0WPbkFEvOPHPbWOYwUAtDQpqZwZkDzJI9DGzFuLQYHMhJkOZtu504g5mBTfpPwmEbgxsJDNJAvyQ24NJy/k3yb962c/UgLUBbwFqYf+PVciP9GNCWAwkSUC1mQC1s+G258YbNcmZbsuEMZsZmC4a0NGPJMwfbLHsM8PjlfPrjGx/b7OQl+A8fvMFgY2PYd7z58I0fFcdwhhgDAw8sJhghLlE4AGIYHEvArYX9ASpfHqKzBo+WUTAKRsEoGGEAALVlUmtpFryyAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0003-4314-5093","institution":"Earthquake Research Institute, The University of Tokyo","correspondingAuthor":true,"prefix":"","firstName":"Hiromichi","middleName":"","lastName":"Nagao","suffix":""},{"id":543657654,"identity":"b2cb6cc6-663a-45b2-a760-85185b4aa3e7","order_by":2,"name":"Shinya Katoh","email":"","orcid":"https://orcid.org/0000-0003-3474-9354","institution":"Earthquake Research Institute, The University of Tokyo","correspondingAuthor":false,"prefix":"","firstName":"Shinya","middleName":"","lastName":"Katoh","suffix":""},{"id":543659203,"identity":"112771a6-be7a-4d2b-b84c-64892aae332b","order_by":3,"name":"Masanao Shinohara","email":"","orcid":"https://orcid.org/0000-0003-1327-509X","institution":"Earthquake Research Institute, The University of Tokyo","correspondingAuthor":false,"prefix":"","firstName":"Masanao","middleName":"","lastName":"Shinohara","suffix":""}],"badges":[],"createdAt":"2025-11-11 00:09:35","currentVersionCode":2,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-8080988/v2","doiUrl":"https://doi.org/10.21203/rs.3.rs-8080988/v2","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":96709960,"identity":"8cee14a8-1cc2-414e-82c8-8f3de136adc2","added_by":"auto","created_at":"2025-11-25 10:09:50","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10438542,"visible":true,"origin":"","legend":"","description":"","filename":"MendoPerezetalSVRDASpreprint.docx","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/0f014f1dd75caefb5fc53835.docx"},{"id":96659841,"identity":"a1135e6c-55e0-419a-a7f1-6f3894d7f71e","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs8080988.json","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/3d95e4c5cad7f71e9920abdf.json"},{"id":96710695,"identity":"f697e211-4fcc-4b60-9029-1bf53e823259","added_by":"auto","created_at":"2025-11-25 10:11:05","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":150528,"visible":true,"origin":"","legend":"","description":"","filename":"rs80809880enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/f57c52aef14fad7632cd426c.xml"},{"id":96709857,"identity":"0f005114-6e55-4157-8807-05a5e4241fd1","added_by":"auto","created_at":"2025-11-25 10:09:44","extension":"emf","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4282460,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.emf","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/f573775f4aa184ec55ebd811.emf"},{"id":96710189,"identity":"fb4dd178-0934-4cd3-8a87-16edbd7fff02","added_by":"auto","created_at":"2025-11-25 10:10:16","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":353416,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/15e174526f1ba21953222e6a.png"},{"id":96709863,"identity":"5e1f1cbc-4de5-4a95-a939-26bb2a629454","added_by":"auto","created_at":"2025-11-25 10:09:44","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1644948,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage11.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/5ea69af5c16602ed6770cf64.jpeg"},{"id":96659839,"identity":"967814a6-d4ba-4b4a-b3be-48fafdb8cb87","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":159229,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/0e612a674e50768ffcc5f208.png"},{"id":96659855,"identity":"e96a7ef9-ee7b-4dcb-8e35-9a31391031b6","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":563180,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/3d2f32d6a0b2412a2f2759cc.png"},{"id":96709515,"identity":"a3168212-6c93-4ce3-a5ca-434c3c37185c","added_by":"auto","created_at":"2025-11-25 10:09:10","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":186159,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/5283b5597ce197a3ee6b2d32.png"},{"id":96709661,"identity":"045eb08b-2195-4efb-8179-6c0bd4448b1f","added_by":"auto","created_at":"2025-11-25 10:09:29","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1157669,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/1ce5a2f7fd21c8d9ad9f8d68.png"},{"id":96710451,"identity":"551481fd-bd17-4a66-a8ae-e6922d5f9fb1","added_by":"auto","created_at":"2025-11-25 10:10:41","extension":"jpeg","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":500096,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/a0f3039c4f51f0abfd42f9e1.jpeg"},{"id":96659876,"identity":"2b1e95f6-77ce-4ac3-8d15-57efed5578a2","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1087969,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/80f5c55efdf0939f17093dfd.png"},{"id":96659868,"identity":"5e2c2400-d250-4421-a72f-d18f5d5d304e","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":364445,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/ef8b7675d0b9587519c9cf89.png"},{"id":96709658,"identity":"d154334f-8051-4d71-afba-72c2a970fb8a","added_by":"auto","created_at":"2025-11-25 10:09:29","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2219968,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/04ba279d2c21e0bb49463e06.png"},{"id":96710769,"identity":"c0e5b4cf-f873-4d43-90e5-31b38d644831","added_by":"auto","created_at":"2025-11-25 10:11:10","extension":"emf","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1050804,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.emf","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/172c17d660381395b91a5993.emf"},{"id":96710083,"identity":"1acf60b8-945c-4daa-8637-95e61b468035","added_by":"auto","created_at":"2025-11-25 10:10:03","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":77757,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/a354d4e3fa7c34e6fd5238b5.png"},{"id":96659861,"identity":"4f3e0a4d-a20e-4c7e-86fc-86d74cc67719","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":67482,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/442bb36bdaef83f2dc54329f.png"},{"id":96659866,"identity":"bb1cb3f4-6c6f-4797-baea-e1775236e00a","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":507731,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/f3703ded6f09ea22e557b6ef.png"},{"id":96659857,"identity":"f762014b-8819-41de-bc31-f6ce69d43d72","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40894,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/ea0dd441fb45231a49b2b1a1.png"},{"id":96659862,"identity":"208eee7b-5e46-4956-928d-ac46cb05509c","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":86583,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/9d9a3e6890a317b1e8ecfc1c.png"},{"id":96659869,"identity":"2d6a0486-87ae-43c9-8a26-653431894b34","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":47297,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/a1942597be48a22ea2183a61.png"},{"id":96659865,"identity":"5ca2814f-a0b5-4414-8fed-859becb28138","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":140189,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/a4b05a5be320861487946abf.png"},{"id":96659870,"identity":"3376a12b-9d14-4004-b068-e51df7513063","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":91836,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/e84da1782542b57e3ca1c2d0.png"},{"id":96709963,"identity":"52c19756-3b41-450c-b603-5bac5c1a6f66","added_by":"auto","created_at":"2025-11-25 10:09:50","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":145207,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/4ae75ae423adcac1bae2758c.png"},{"id":96709655,"identity":"317369fc-5e08-471f-93ee-dc8bf9823b54","added_by":"auto","created_at":"2025-11-25 10:09:28","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":94895,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/9f20640d0718cb9425e3df0d.png"},{"id":96659864,"identity":"c8b92640-be90-4058-a2f1-98d969e1d227","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":298543,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/24ab9f056b11ef69cde3f3f0.png"},{"id":96710025,"identity":"04c6b274-2117-46a1-b157-b0a546edf23c","added_by":"auto","created_at":"2025-11-25 10:09:54","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6232,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/53cb9522e0c64db4892e43a6.png"},{"id":96709912,"identity":"bad8d269-dc4e-48ab-b5dd-56df4b202414","added_by":"auto","created_at":"2025-11-25 10:09:47","extension":"xml","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":148605,"visible":true,"origin":"","legend":"","description":"","filename":"rs80809880structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/a9f8a8c02aae39df7948975e.xml"},{"id":96710484,"identity":"1a74496e-4819-4c66-b1df-21de762e72a0","added_by":"auto","created_at":"2025-11-25 10:10:43","extension":"html","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":156001,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/599f789b4ca015a194e4e58d.html"},{"id":96709831,"identity":"789c4dc2-57b3-4512-9fa4-a577bf9cc3c8","added_by":"auto","created_at":"2025-11-25 10:09:43","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":3009860,"visible":true,"origin":"","legend":"\u003cp\u003eMap of the offshore section of Sanriku, Japan. The black line shows the relative position of the seafloor observation cable in Sanriku, and the blue line indicates the cable section that is buried beneath the sediment layer. The red inverted triangles indicate the position of the ocean-bottom accelerometers of the Sanriku seafloor observatory system.\u003c/p\u003e","description":"","filename":"Fig1.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/982b4199db7aaa95b558a4dc.png"},{"id":96659835,"identity":"5d1b3609-d42f-44a8-a715-c3e7736332e4","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1468407,"visible":true,"origin":"","legend":"\u003cp\u003ea) Flow diagram of the method followed to detect earthquake signals in DAS data. b) Example of a template waveform (upper panel), DAS traces (middle panel), and correlation coefficients (CC) (lower panel) obtained after the application of template matching. Both template and DAS data were recorded on 2023-07-26 12:26:03 UTC.\u003c/p\u003e","description":"","filename":"Fig2.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/efe8c262770eaa187286f3f5.png"},{"id":96710158,"identity":"a7c6f4b4-3079-47d6-b4ec-b7944176ed64","added_by":"auto","created_at":"2025-11-25 10:10:11","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":526478,"visible":true,"origin":"","legend":"\u003cp\u003ea) Flow diagram of the method used to pick P and S phases on the DAS traces. b) Examples of two bandpass filtered DAS traces and their associated P and S picks. The red and blue lines correspond to the P and S phases, respectively. The two traces belong to the DAS data from 2023-07-26 12:26:03 UTC.\u003c/p\u003e","description":"","filename":"Fig3.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/db4246ee16d552e103223ed6.png"},{"id":96659838,"identity":"61652c66-0397-4e6a-8ea1-eb65cf865b0f","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":3199933,"visible":true,"origin":"","legend":"\u003cp\u003eExample of the obtained P and S picks from the 2023-07-26 12:26:03 UTC DAS data. The figure shows the a) initial picks obtained after picking comparison, b) pickings after obtaining the initial SVR models, and c) refined models after applying spectrogram-based picker.\u003c/p\u003e","description":"","filename":"Fig4.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/d2593f3dc1788b0298c78361.png"},{"id":96709822,"identity":"1ed618af-ff37-4fbd-8a59-9de40900b418","added_by":"auto","created_at":"2025-11-25 10:09:43","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":282480,"visible":true,"origin":"","legend":"\u003cp\u003eHistograms (upper panels) and scatter plots (lower panels) of the difference of P and S arrival times of manual (black points) and SVR-DAS (red points) picks.\u003c/p\u003e","description":"","filename":"Fig5.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/2302c79ac0bcf2d8f09afa4b.png"},{"id":96709581,"identity":"860a2169-690e-44fb-95a8-4a5f98425b12","added_by":"auto","created_at":"2025-11-25 10:09:19","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":2499079,"visible":true,"origin":"","legend":"\u003cp\u003eNumber of detections per minute obtained by applying envelope template matching from 2022-02-28 08:51:20 to 2022-02-28 23:59:20 UTC. The red line shows the cumulative frequency of earthquakes. The numbers are associated with the DAS traces (left column) and the CC plots (right column).\u003c/p\u003e","description":"","filename":"Fig6.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/938988bb014d943ac941268a.png"},{"id":96710752,"identity":"55c0b612-c5af-4023-82af-2d0443dda4f7","added_by":"auto","created_at":"2025-11-25 10:11:09","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":964078,"visible":true,"origin":"","legend":"\u003cp\u003eExamples of initial picks obtained from DAS traces from a) 2022-02-28 10:16:20, b) 2022-02-28 10:31:20, and c) 2022-02-28 12:01:20 UTC. The left column shows examples of good picks agreement, and the right column shows examples of disagreement between picks.\u003c/p\u003e","description":"","filename":"Fig7.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/035fba67049758a99066bd82.png"},{"id":96709431,"identity":"27ec52be-02ee-4d24-bf9e-cd75a2cd930e","added_by":"auto","created_at":"2025-11-25 10:09:00","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":4154788,"visible":true,"origin":"","legend":"\u003cp\u003eExamples of initial picks (left column) and refined models (right column) obtained from DAS traces from a) 2022-02-28 10:16:20, b) 2022-02-28 10:31:20, and c) 2022-02-28 12:01:20 UTC.\u003c/p\u003e","description":"","filename":"Fig8.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/2854bab39ef7595dcd355787.png"},{"id":96710647,"identity":"9549067d-6953-4ea1-b2e3-4b907feaa9bf","added_by":"auto","created_at":"2025-11-25 10:11:02","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":205052,"visible":true,"origin":"","legend":"\u003cp\u003ea) Loss and b) accuracy curves from training (blue line) and validation (red line) datasets obtained after training the CNN-RNN model.\u003c/p\u003e","description":"","filename":"Fig9.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/c91b648719678eafab610361.png"},{"id":96659849,"identity":"232839a4-86fd-4ca7-8c5e-fc7241e0df0d","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":435910,"visible":true,"origin":"","legend":"\u003cp\u003eExamples of observed (solid line) and predicted (dashed line) labels from the validation datasets of DAS data. The red line corresponds to the P label, the blue line to S label, and the green line to the noise label.\u003c/p\u003e","description":"","filename":"Fig10.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/814fd89817a842eafed7f2c7.png"},{"id":96659863,"identity":"afdcd9cf-f8d2-4b5d-a3e5-c118ecfe7a0d","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":4414755,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of P and S picks obtained from PhaseNet-DAS (Zhu et al., 2023) (left column) and SVR-DAS (right column) for a) 2022-02-28 10:16:20 h, b) 2022-02-28 12:01:20 h c) 2023-07-26 12:26:03 h events.\u003c/p\u003e","description":"","filename":"Fig11.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/48c4ef415f1e3487b14bb4b4.png"},{"id":96659867,"identity":"0e3cbd16-11db-47b4-a79d-dfdce82a0d55","added_by":"auto","created_at":"2025-11-24 18:04:16","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":259090,"visible":true,"origin":"","legend":"\u003cp\u003eHistograms of the time arrival difference between methods. The superscript x indicates to either PhaseNet-DAS (blue bars) or manual picking (red bars). The histograms are associated with a) 2022-02-28 10:16:20 h, b) 2022-02-28 12:01:20 h, and c) 2023-07-26 12:26:03 h.\u003c/p\u003e","description":"","filename":"Fig12.png","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/a461a433257a3189cd88e50c.png"},{"id":96913086,"identity":"f106aac1-3351-4826-b978-ef75a82f4500","added_by":"auto","created_at":"2025-11-27 13:51:56","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":20507790,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/a3fdf476-f67d-43c9-a93b-57c674b53db5.pdf"},{"id":96709722,"identity":"68364b97-7bb0-43c2-8f35-5bac1d229d26","added_by":"auto","created_at":"2025-11-25 10:09:34","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":9079842,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Information\u003c/p\u003e","description":"","filename":"MendoPerezetalDASSuppFile.docx","url":"https://assets-eu.researchsquare.com/files/rs-8080988/v2/430106ad2140e6fce4abbdfa.docx"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"SVR-DAS: A Machine-Learning-Based Method to Create Earthquake Catalogs from Seafloor Distributed Acoustic Sensing Measurements","fulltext":[{"header":"Introduction","content":"\u003cp\u003eMonitoring offshore earthquake activity is an important research field for seismic hazard assessment. Most of the large megathrust source areas are located offshore near subduction zones. For this reason, marine earthquake monitoring systems have been installed over the last 20 years, allowing us to obtain offshore seismic and geodesic observations. In Japan, seafloor observatories such as the Dense Ocean Floor Network System (DONET) (Kawaguchi et al., \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2011\u003c/span\u003e; Aoi et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), N-net (Aoi et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2023\u003c/span\u003e)d net (Kanazawa et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Aoi et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) are currently operating in Nankai and Japan Trench regions, respectively.\u003c/p\u003e\u003cp\u003eAlthough several advances have been achieved by current marine infrastructure, establishing dense marine observation networks remains a challenge due to the difficulty to install seafloor observatories in deeper areas and recovering data. In addition, the maintenance and lifespan of seafloor seismometers is very limited, and this can result in extremely costly observation systems.\u003c/p\u003e\u003cp\u003eAn emergent solution to the observational gaps in deep marine areas is the use of Distributed Acoustic Sensing (DAS) measurements, which is transforming earthquake monitoring. DAS measures the phase changes between an induced laser pulse and Rayleigh backscattering produced by inhomogeneities in the fiber structure (Lindsey and Martin, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). This phase change is linearly proportional to the strain of fiber optic cables to earthquakes or any other external source that perturbs the cable. The interrogator unit (IU) sends controlled laser pulses at a specific wavelength (near 1550 nm) throughout the optical fiber, and a phase change of the scattered light is measured by the IU (Masoudi and Newson, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Hartog, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Lindsey and Martin, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Zhan, 2020). The strain rate is proportional to the change in the gauge length of scattered light obtained by emitting laser pulses. Hence, DAS enables fiber optic cables to act as linear strainmeters (Benioff, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e1935\u003c/span\u003e), allowing the retrieval of earthquake signals (Shinohara et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; 2021; Lindsey and Martin, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eTwo main advantages arise from the use of seafloor optic fiber to monitor earthquakes. First, DAS measurements are equivalent to a linear array of thousands of channels, allowing to record earthquakes with unprecedented resolution (Lindsey and Martin, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). The second is that DAS measurements can be done using spare (dark) fibers from already installed telecommunication cables. DAS was first applied in oil and gas industry, where encased fiber optic cables in boreholes were used for seismic exploration (Daley et al. \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). In seismology, teleseismic waves have been retrieved from broadband DAS measurements (Lindsey et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Yu et al., \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eDespite its promise, many challenges remain to be overcome. One is the manageability of high volumes of data. The size of one day of continuous DAS data may rise to several tens of TB (depending on the sampling rate and the gauge length). Consequently, large computing clusters are required for storage, which may be insufficient for real-time applications. Also, it is impossible for analysts to process high-density data in a reasonable time frame. Another issue is related to the cable coupling to the ground. Several laboratory experiments suggest that loosely installed cables produce weaker responses (Papp et al., 2017; Becker et al., 2018; Lindsey and Martin, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Ajo-Franklin et al., 2021). In field observations, urban activity often produces stronger signals than earthquakes (Lindsey and Martin, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), and in colder areas, the sensitivity to high-frequency sources is variable in the presence of snow and ice (Castongia et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Walter et al., \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Lindsey and Martin, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eIn recent years, deep learning has emerged as a new paradigm in science for data processing due to its ability to make data-driven decisions based on nonlinear functions. In seismology, well-established methods have been developed for earthquake detection and phase picking (Ross et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Zhu and Beroza, \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Mousavi et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Kaneko et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Tokuda and Nagao, \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Katoh et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), seismic phase association (Ross et al., \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Zhu et al., \u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; McBrearty and Beroza, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), polarity determination and focal mechanism estimation (Zhang et al., \u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Katoh et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), denoising (Yu et al., \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), waveform inversion (Liu et al., \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), and modeling (Moseley et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). For DAS data, there are already proposed methods for earthquake detection and seismic picking yielding accurate results (Zhu et al., \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Ding et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eCurrent deep learning models used for DAS data are trained using inland seismic and DAS records. While these models successfully capture broad earthquake-related features related to the earthquake signal, finer details associated with noise present in seafloor environments are not being considered, leading to errors in seismic phase picking for some events. To the best of our knowledge, the seafloor DAS-based training datasets of seafloor optic fiber cables are still scarce and currently is an active research field (Xiao and Tillman, 2024). In this work, we address this obstacle by creating training/validation datasets from seafloor DAS data that can be used to develop and/or test deep learning models. Our framework consists of two stages: earthquake detection and body wave phase picking (P and S). In the detection stage, we apply template matching using envelope waveforms to identify earthquake signals in DAS data. We begin with one template retrieved from DAS data, and all subsequent detections are used to expand the template library. In the phase picking stage, we apply recursive STA/LTA (Allen, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1982\u003c/span\u003e) and autoregressive models using the Akaike Information Criterion (AR-AIC) (Akazawa et al., 2004) to obtain P and S picks. We compare the obtained picks, and the consistent ones are selected as training datasets to create P and S pick models. To create these models, we apply Support Vector Regression (SVR) (Vapnik, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e1997\u003c/span\u003e), a machine-learning method that searches for hyperplanes that best fits the data. First, we test the method using well-defined DAS data from an earthquake that occurred near the cable on 2023-07-26 12:26:00 UTC. Afterwards, we apply SVR-DAS on DAS data from 2022-02-28. Finally, we train and validate a CNN-RNN model inspired by trigger-word systems (Supriya et al., \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). The obtained picks from SVR-DAS are compared with picks from PhaseNet-DAS (Zhu et al., \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), a deep-learning-based method for earthquake detection and seismic picking in DAS data. We believe that SVR-DAS method will provide accurate arrival-times estimates from seafloor DAS data.\u003c/p\u003e"},{"header":"Data and Methods","content":"\u003cp\u003eThis work focuses on the detection of earthquake signals and the identification of P and S phases in seafloor DAS measurements. We employ the data obtained from the Sanriku seafloor observation system located near Iwate prefecture, in northeastern Japan (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). This observatory is being operated by the Earthquake Research Institute, The University of Tokyo (Shinohara et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). The system consists of three three-component accelerometers and two tsunami-meters connected to a 120-km optic fiber cable. Two cables are installed in the observation system: the first one was installed in 1996 and performs data transmission and system control, and the second was installed in 2015. The cable installed in 1996 has a total of 12 optic fibers. Six of these are used for data transmission, and the rest of them are spare (dark) fibers, which are suitable for DAS measurements (Shinohara et al., \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). In addition, because there are no signal repeaters through the system the DAS measurements can be performed to the end of the cable. Both cables are connected to a land station in Kamaishi, Japan (Shinohara et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; 2021; \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eWe propose a framework, called SVR-DAS, that performs earthquake detection and seismic phase picking. To detect earthquake signals, we apply the template matching method (Gibbons and Ringdal, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2006\u003c/span\u003e). Template matching measures the similarity between the target trace and a master signal (or template) using the correlation coefficient (CC). This method has been used for both inland (Li and Zhan, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) and seafloor DAS data (Miao et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Here, we use a different implementation of template matching. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea shows the workflow of our implementation. First, we use envelopes instead of waveforms for the search. The envelope shape is simpler than the waveform, which enhances the CC values. Second, we use an initial template to perform the search. The initial template is retrieved from the DAS data itself, and the detections from the first search are used as templates for subsequent searches. The search is performed using 20-s time windows with 50% overlap. Both template and DAS waveforms are band-pass filtered between 5 and 15 Hz to reduce ocean noise and suppress high frequency signals.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb shows an example of the results from applying template matching with one template. The data corresponds to an earthquake from 2023-07-26 12:26:00 UTC (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb, middle panel). The template used is from channel 800 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb, upper panel), located about 32 km from the landing station. This channel corresponds to the buried section of the cable. Along the cable, we observe a clear pattern of CC values above 0.8 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec, middle and lower panels) that coincides with the time at which the earthquake signal appears in the DAS data. In the rest of this section, we present additional results obtained from the DAS data of this event.\u003c/p\u003e\u003cp\u003eFalse detections may arise due to the similarity between the envelope and the later sections of DAS traces. However, we can verify whether the DAS data contains a signal by counting the number of false detections per file. The number of false detections caused by noise is relatively constant. Hence, the DAS data is considered to contain signals if we observe an increase in the number of detections above a mean value. Therefore, we set a detection threshold based on the mean number of false detections in DAS data. Through trial and error, we set this threshold to 30 counts.\u003c/p\u003e\u003cp\u003eAfter detection, the algorithm proceeds to identify P and S arrival times. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea shows the flow diagram that summarizes the seismic phase picking phase performed by SVR-DAS. First, we load two files: the file that contains all detections obtained by template matching, and the original file. The data is once again band-pass filtered between 5\u0026ndash;15 Hz due to the algorithm stores unfiltered detections. In addition, we increase the time length of the detections from 20 to 30 s and estimate the SNR of all detections. We assume that the P wave may be missed within the 20-s time window if the SNR is relatively low. Hence, we change to 30-s time window length. Otherwise, P phase may be within the time windows, and the algorithm use the 20-s time windows instead.\u003c/p\u003e\u003cp\u003eSVR-DAS applies two picking methods to identify P and S in the DAS data: recursive STA/LTA (Allen, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1982\u003c/span\u003e), and auto regressive models using the Akaike Information Criterion (AR - AIC) (Akazawa, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2004\u003c/span\u003e). Each method has its own fixed parameters, and these are summarized in Table S1 from Supplemental Material. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb shows two DAS traces obtained from 2023-07-26 data and their corresponding P and S picks obtained using STA/LTA and AR-AIC. Good agreement can be observed in P and S picks obtained from the two methods.\u003c/p\u003e\u003cp\u003ePick disagreements between methods may arise due to the variable SNR between traces. Hence, the picks obtained from these methods are compared with each other, and a pick coherency threshold is set to discriminate between coherent and non-coherent picking. By coherent picking, we refer to the agreement between pick estimations. We set the pick threshold to 3 s, and the initial pick is set as the mean both picks if the time difference between picks is below the threshold. On the contrary, if the separation between peaks is above 3 s, then the algorithm discards the picks.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe pick coherency threshold should filter most of the incoherent picks. However, it is still plausible the case where all methods agreed in a wrong pick estimation. For this reason, after obtaining initial picks, the algorithm performs an outlier removal process that consists of the evaluation of the slope changes between adjacent picks obtained from the DAS traces. We take advantage of the spatial information of DAS data, assuming that coherent picking between channels must have minimum slope changes. If no outliers are present, then the slope change between picks is constant through all points. In contrast, the outliers may change abruptly the slope. The algorithm estimates the slope for all obtained picks and discards the pickings that separate from the mean slope. After applying this procedure 2023-07-26 DAS data, we observe that most of the obtained P and S picks align agrees with the apparent arrival times observed in the DAS data (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eUp to this point we may obtain reasonable estimations of P and S arrival times. To further improve the estimation of P and S picks we apply Support Vector Regression (SVR), a regression method based on Support Vector Machines (SVM). SVM is a supervised Machine Learning (ML) method used for classification, regression and outlier detection, and consists of the search of hyperplanes in higher dimensional spaces that best separates the data (Vapnik, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e1997\u003c/span\u003e; Pedregosa et al., 2012). The application of SVM in regression problems is the SVR method, which searches best hyperplanes that fits the data. Linear and nonlinear regression is possible due to kernel search instead of model parameters search (Smola and Scholkopf, 2003; Pedregosa et al., 2012). An advantage of SVR is that only a subset of training dataset is needed to perform the regression. For more details about the theoretical aspects of SVM and SVR, the reader can refer to Smola and Scholkopf (2003).\u003c/p\u003e\u003cp\u003eWe implement SVR algorithm provided by the Python package Scikit-Learn (Pedregosa et al., 2012). The kernel selection is based on two parameters: the density of P and S picks through the data, and the interquartile range (IQR) of P and S picks. By trial and error, we use linear regression to the data with IQR above than 2.1. In addition, the linear regression is also set to the data where the density of picks is concentrated in a small section of DAS data. Otherwise, we set Radial Basis Functions (RBF) as the kernel to perform the regression. The reason for this choice is that data with small IQR may be associated with events that occurred near the cable. Thus, the P and S picks behavior may present a nonlinear trend.\u003c/p\u003e\u003cp\u003eP-to-S conversions are observed in ocean-bottom seismometers and DAS measurements due to the sediment layers in the Sanriku area (Fujie et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Fukushima et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Because DAS is prone to record strain in horizontal direction due to the aspect ratio of the optic fiber, it is most likely that the amplitude of the P-to-S converted wave is more prominent than P waves. Assuming that the model for P may be related P-to-S converted waves, we need to refine our picking to capture as accurately as possible the P wave arrival time.\u003c/p\u003e\u003cp\u003eHence, we adopt a procedure based on the signal spectrogram to refine the P picks obtained from the SVR model. The spectrogram is calculated in a 20-s time window centered in the estimated P pick. If there are no other impulsive signal or noise inside the time windows, it is probable that we observe in the spectrogram a change in the power of the signal due to the P wave. Then, we create envelope functions using the obtained spectrograms. The envelope functions are normalized with respect to the maximum value of the spectrogram. We also attempt to refine S wave picking using this approach if there is a moveout due to the SVR regression. An amplitude threshold is set to retrieve both refined P and S arrival times. For P wave, the threshold is set as the 20% of the median absolute deviation (MAD) of the envelope function. Because the P wave amplitude decreases at the relatively far section of the cable, the threshold changes to 10% of the MAD value. For S wave, the threshold is set as the 80% of the MAD of the envelope. An example of this envelope function, and the pickings obtained using this procedure is shown in Figure S1 from Supplemental Material.\u003c/p\u003e\u003cp\u003eMost of the picking is automatic and self-guided by the algorithm. Nevertheless, a manual revision is made for all pickings. At this point, P and/or S pickings may be discarded or readjusted. The final catalog is stored in an HDF5 file following the STEAD database format proposed in Mousavi et al. (\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eWe evaluate the performance of the method by comparing the SVR-DAS pickings with manual picks obtained of 400 channels from DAS data from 2023-07-26. These channels lie within the first 80-km DAS cable section (see Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The evaluation of the picks is summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows the histogram and the comparison between SVR-DAS and manual picks. We compare both datasets by estimating the standard deviation (σ) and the IQR of P and S picks. The manual choice of P and S arrival times across the channels must be lower than the obtained from picking algorithms. The difference between manual picking and SVR-DAS of IQR and σ is around 0.1 and 0.6 except for the IQR of P picks, which increases to 1.92. The difference of arrival times between manual and SVR-DAS for most of P and S picks is within \u0026plusmn;\u0026thinsp;1 s. For P, we observe that the picks distribution roughly coincides up to 40 km. Above this point, the pick distributions associated with manual picks are more widespread. This is reasonable due to the decrease in SNR for the farthest channels. In contrast, the distributions of S picks do not vary significantly for most of the data. It seems that there is less uncertainty in the choice of S wave arrival times. However, manual picking was hard for most of the channels that correspond to the cable section above 40 km.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of the standard deviation (σ) and interquartile range (IQR) between manual and SVR-DAS picks.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eValue\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eσ\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eIQR\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003et\u003csup\u003eSVR\u003c/sup\u003e\u003csub\u003eP\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2.63\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e5.02\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003et\u003csup\u003eM\u003c/sup\u003e\u003csub\u003eP\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2.22\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e3.77\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003et\u003csup\u003eSVR\u003c/sup\u003e\u003csub\u003eS\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e3.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e3.69\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003et\u003csup\u003eM\u003c/sup\u003e\u003csub\u003eS\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e6.69\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e6.59\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e(t\u003csub\u003eS\u003c/sub\u003e - t\u003csub\u003ep\u003c/sub\u003e)\u003csup\u003eSVR\u003c/sup\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.36\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e1.94\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e(t\u003csub\u003eS\u003c/sub\u003e - t\u003csub\u003ep\u003c/sub\u003e)\u003csup\u003eM\u003c/sup\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1.66\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e2.43\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eWe applied SVR-DAS to the data from 2022-02-28 08:51:20 to 2022-02-28 23:59:20 h. All data is in UTC format. Figure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e shows the results obtained from template matching algorithm. The data was acquired using a gauge length of 100 m, and the spatial sampling is 16.44 m. The acquisition sampling rate was set to 800 Hz; however, we decimated the DAS data to 100 Hz. We set as the initial template the same DAS trace from 2023-07-26 12:26:02 h (see Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The algorithm identified 23 events with mean SNR values ranging from 0.03 dB to 0.27 dB (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ea). The CC values of all detections range between 0.8 and 0.9. The number of detections varies between 3 and 870, with a mean value of 25 detections per DAS file.\u003c/p\u003e\u003cp\u003eWe plotted the DAS data and their associated CC values of three identified events (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eb). For plotting, we increased the window size of the DAS data to 180 s to avoid incomplete events. The blue dotted line in each panel of DAS data highlights the original file size used in the template matching search. We observe that the highest values of CC coefficients coincide with the sections of DAS data where the highest amplitudes are present. The values that coincide with the DAS data at 60 s are discarded when retrieving the detections from the data.\u003c/p\u003e\u003cp\u003eNot all detections are associated with earthquakes. We can identify the unrelated signals and separate them from the earthquake signals using the standard deviation of the CCs. From the 23 detections obtained from template matching, 8 of these detections contained earthquake-related strain rate signals with arrival time differences S \u0026ndash; P above 20 s, and 3 are pure noise. From the remaining events, the algorithm estimated P and S picks from 8 events, and the other 4 most of the pickings were from S waves. We build the training and validation dataset using the P and S pickings from these 8 events.\u003c/p\u003e\u003cp\u003eA comparison between initial picks obtained from STA/LTA and AR-AIC methods is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e. We show six DAS traces from 2022-02-28 10:16:20 UTC (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003ea), 2022-02-28 10:31:20 UTC (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eb) and 2022-02-28 12:01:20 UTC (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003ec). In the left column, we show the traces where there is well agreement between the picks, and in the right column we show the traces with picks that do not fulfill the pick coherency threshold criterion. For relatively high SNR signals both STA/LTA have good agreement. AR-AIC method is more robust to the presence of noise than STA/LTA. However, mispicking may arise in data with non-earthquake impulsive signals (K\u0026uuml;perkoch et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2011\u003c/span\u003e). For non-coherent picks, we observe that the difference between obtained picks for P, S picks or both ranges from 4 to 15 s. This range may increase in other channels if the SNR is relatively lower (see Figures S2-S3 from Supplemental Material).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe training points for SVR and the obtained pick models for P and S are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e. The number of training examples is not fixed and varies with the available number of data points. In addition, the parameter for obtaining the model differs depending on the distribution of points and the SNR. For the highest SNR examples, we use an RBF kernel with regularization parameter C of 1.0 and an epsilon-tube parameter ε of 0.1. After several trial-and-error tests, we set these parameters which are the default values for this method (Pedregosa et al. 2011). However, if we use a linear kernel we decrease both C and ε parameters to avoid noncoherent solutions. Thus, we set C to 0.1 and 0.05 for linear fitting for P and S models, respectively. The pick models obtained from SVR are consistent with the observed P and S pattern in the DAS data. Figure S4 from Supplemental Material shows another example of a detected event with their associated P and S picking models.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAlthough, we can observe a good fit between the results and the observed data some points must be considered when using SVR. First, the models may be unstable with low-SNR data. We observe in the P and S models from 2022-02-28 12:01:20 UTC a clear deviation of the P model starting after 40 km. This observation coincides with an increase in noise in the DAS data. A better fit may be achieved if we decrease the C parameter from the SVR model. However, it may force the model to fit a line, producing a non-realistic fitting. Therefore, we could see more model misfits at the channels with lower SNR.\u003c/p\u003e\u003cp\u003eThe obtained results from SVR-DAS are then used as training and validation datasets to train a deep learning model. For this purpose, we use a CNN-LSTM-based model (Mendo-P\u0026eacute;rez et al., \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) for sequential data. The architecture of this model is based on the ones used in trigger voice systems (Supriya et al., \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). As the model input, we use the spectrogram of the individual DAS traces, calculated using 0.6-s time windows with an 95% overlap. Prior to create the spectrograms, the data is downsampled from 100 Hz to 33.3 Hz. In addition, the data is normalized by subtracting the mean and dividing it to the standard deviation.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo train this model, we use a subset of 18000 DAS data labels from our catalog. From these, 12000 examples are associated with earthquake signals, and 6000 examples are noise. The subset is separated into training and validation datasets using a ratio of 8:2. We use the Mean Squared Loss, as the loss function, and for backpropagation stage we use the AdamW optimizer using a fixed learning rate of 0.001. We apply a 4-s width Gaussian masks to create the output labels. The maximum probability of the masks is set to 1.0 at the time position of each P and S arrival time obtained from SVRDAS. The model is trained using 20 epochs and a batch size of 8. The model took around 1 hour to complete training using an NVIDIA GeForce RTX 3060. Figure\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e shows both the loss and accuracy curves. Here, we define accuracy of the model as the ratio of the number of labels correctly obtained by the model and the total number of labels. Assuming the accuracy is 100% if the model correctly identifies all examples, the accuracy of the model at the end of the training stage was 96% for both training and validation datasets.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe obtained labels from the CNN-LSTM model can be seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e. Here, we show four random DAS traces and their associated spectrograms retrieved from the validation datasets. We compare the labels obtained from the SVR-DAS method and the labels obtained by the CNN-RNN model assuming that the SVR-DAS labels are our ground truth. The maximum peaks of both P and S labels coincide at the same time as the ground truth labels. The probabilities for P and S label are 0.85\u0026ndash;0.94 and 0.96\u0026ndash;0.97, respectively. We observe that the trained model is prone to identify better the S arrival times than P. This problem may be resolved by adding more training examples of well-defined P arrival times in DAS traces with variable SNR.\u003c/p\u003e\u003cp\u003eWe also compare the arrival times obtained from SVR-DAS to the ones obtained by PhaseNet-DAS (Zhu et al., \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), a deep-learning-based method for seismic phase picking, to estimate P and S arrival times. This model follows a semi-supervised scheme by using a pretrained PhaseNet model (Zhu and Beroza, \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), to estimate initial DAS picks, followed by phase association using GaMMA (Zhu et al., \u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) to filter non-coherent picks. The obtained picks are used to train a UNet-based model similar to PhaseNet (Rommberg et al., 2015) to estimate P and S arrival times in DAS data.\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e shows the picks obtained from three events: 2022-02-28 10:16:20 h (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003ea), 2022-02-28 12:01:20 h (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003eb), and 2023-07-26 12:16:03 h (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003ec). From this point, we will refer to the events as E1, E2, and E3 for simplicity. In these three events, both methods capture most of the arrival times with apparently the same distribution but present some noteworthy differences. Note that PhaseNet \u0026ndash; DAS in E1 also detects a small event between 20\u0026ndash;40 s before to the main event (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003ea). The reason that SVR-DAS did not identify this event is due to the choice of the time windows to run and STA/LTA and AR-AIC pickers. SVR-DAS is thought to identify one event per time window. Hence, a disadvantage in SVR-DAS is that it cannot identify more than one earthquake per time window. To detect consecutive events, smaller time window must be used to fulfill the one event condition. PhaseNet \u0026ndash; DAS estimated P picks up to 80 km coinciding with a decrease in SNR. Using SVR-DAS we can obtain P picks in the low SNR region due to the SVR fitting. In E2 we observe that the P pick estimation stops near 100 km, where the noise starts to increase. In this case, SVR-DAS also provide smoother results. The S pick distributions of both methods in both events coincide very well. Finally, in E3, we observe that there is a small gap in PhaseNet-DAS P picks between 60 and 80 km. Another feature is the distribution of P picks around 20 s with a slope nearly horizontal. A closer inspection to each one of the DAS traces was done shows that this feature seems related to a non-physical event. In contrast, SVR-DAS P picks show a continuous distribution with no gaps.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo quantitatively evaluate the picking performance of SVR-DAS, we estimate the time arrival difference obtained between SVR-DAS and both PhaseNet-DAS (t\u003csup\u003eSVR\u003c/sup\u003e - t\u003csup\u003ePN\u003c/sup\u003e) and manual picking (t\u003csup\u003eSVR\u003c/sup\u003e \u0026ndash; t\u003csup\u003eM\u003c/sup\u003e) for P and S phases. The histograms associated with these quantities are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e. In addition, Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the mean (\u003cb\u003e\u0026micro;\u003c/b\u003e), median, minimum (min) and maximum (max) values, standard deviation (\u003cb\u003eσ\u003c/b\u003e), interquartile range (IQR), and Pearson correlation coefficient (R) for the time differences of P and S picks. For simplicity, we use the notation Δt\u003csup\u003ePN\u003c/sup\u003e\u003csub\u003eP,\u003c/sub\u003e Δt\u003csup\u003eM\u003c/sup\u003e\u003csub\u003eP,\u003c/sub\u003e Δt\u003csup\u003ePN\u003c/sup\u003e\u003csub\u003eS\u003c/sub\u003e and Δt\u003csup\u003eM\u003c/sup\u003e\u003csub\u003eS\u003c/sub\u003e to refer to the difference between SVR-DAS and PhaseNet-DAS for P phase, SVR-DAS and manual picking for P phase, SVR-DAS and PhaseNet-DAS for S phases, and SVR-DAS and manual picking for S phase, respectively.\u003c/p\u003e\u003cp\u003eThe distributions for P and S phases associated with manual picking are more widespread than the ones associated with PhaseNet-DAS. According to Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the difference between the standard deviations (Δσ) of P picks distributions associated with PhaseNet-DAS and manual picking for E1, E2 and E3 are 1.08,1.97, and 0.18, respectively. The IQR difference (ΔIQR) for is 0.75 for E1, 0.45 for E2, and 0.21 for E3. In the case of S picks, Δσ is 0.38 for E1, 1.35 for E2, and 0.26 for E3. For ΔIQR, the values of E1, E2 and E3 are 0.14, 0.98, and 0.34. We observe that σ and IQR values for the three events are higher for E1 and E2. Although for E3 is practically the same tendency, the difference is smaller than the observed in the other events.\u003c/p\u003e\u003cp\u003eThe central values of the distributions in E1 and E3 are prone to the left, meaning that PhaseNet-DAS picks and manual arrival times estimations are after SVR-DAS arrival times. In the case of E2, we see that the central value of the P distribution is near 0 s, and the value of S distribution is above 0 s. Another observation is related to the correlation coefficient R between picks. For E1 and E3, we observe R values of PhaseNet-DAS and manual picking is equal or higher than 0.90 for both P and S picks. In contrast, the R values of E2 varies depending on the method. While for PhaseNet-DAS-related picks R is 0.77 s and 0.32 s, for manual picking the values are \u0026minus;\u0026thinsp;0.16 and 0.10.\u003c/p\u003e\u003cp\u003eWith all these elements, we can conclude that picking of SVR-DAS and PhaseNet-DAS is consistent in these events. In contrast, manual picking presents wider distributions due to the bias in the time arrival picking of P and S in relatively lower SNR DAS channels. Particularly, manual picking in E2 was challenging. Although in the DAS data you can clearly see the P and S time arrivals, it is difficult to select accurately in each channel due to the low SNR.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of mean (\u0026micro;), median, minimum value (Min), maximum value (Max), standard deviation (σ), and interquartile range (IQR) estimated from the difference in arrival times of P (Δtp) and S (Δtp) picks obtained from SVR-DAS, PhaseNet-DAS, and manual picking.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"8\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colspan=\"8\" nameend=\"c8\" namest=\"c1\"\u003e\u003cp\u003e2022-02-28 10:16:20 h (E1)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eParameter\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e\u0026micro;\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003emedian\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003emin\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003emax\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003eσ\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003eIQR\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003eR\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔt\u003csup\u003ePN\u003c/sup\u003e\u003csub\u003eP\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-0.52 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-0.54 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-0.92 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.16 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.15 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.99 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔt\u003csup\u003eM\u003c/sup\u003e\u003csub\u003eP\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-1.56 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-1.39 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-6.65 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1.78 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e1.24 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.90 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.90 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔt\u003csup\u003ePN\u003c/sup\u003e\u003csub\u003eS\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-0.46 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-0.43 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-1.17 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.10 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.30 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.44 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.98 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔt\u003csup\u003eM\u003c/sup\u003e\u003csub\u003eS\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-0.53 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-0.48 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-3.70 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e4.33 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.68 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.58 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.99 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"7\" nameend=\"c7\" namest=\"c1\"\u003e\u003cp\u003e2022-02-28 12:01:20 h (E2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔt\u003csup\u003ePN\u003c/sup\u003e\u003csub\u003eP\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-0.12 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-0.08 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-0.75 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.28 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.29 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.51 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.77 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔt\u003csup\u003eM\u003c/sup\u003e\u003csub\u003eP\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-0.33 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.11 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-11.59 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3.26 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e2.26 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.96 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e-0.16 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔt\u003csup\u003ePN\u003c/sup\u003e\u003csub\u003eS\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.45 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.45 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-0.47 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1.51 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.51 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.79 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.39 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔt\u003csup\u003eM\u003c/sup\u003e\u003csub\u003eS\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1.16 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1.43 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-13.72 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e4.37 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e1.86 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e1.77 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.10 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"7\" nameend=\"c7\" namest=\"c1\"\u003e\u003cp\u003e2023-07-26 12:26:03 h (E3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔt\u003csup\u003ePN\u003c/sup\u003e\u003csub\u003eP\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-0.84 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-0.57 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-3.20 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2.16 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.88 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.76 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.98 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔt\u003csup\u003eM\u003c/sup\u003e\u003csub\u003eP\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.40 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.20 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-2.63 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2.66 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.70 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.97 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.96 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔt\u003csup\u003ePN\u003c/sup\u003e\u003csub\u003eS\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-0.26 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-0.31 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-2.25 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.46 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.31 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.31 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.99 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔt\u003csup\u003eM\u003c/sup\u003e\u003csub\u003eS\u003c/sub\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.28 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.15 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-2.14 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2.71 s\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.57 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.65 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.98 s\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eHere, we present a ML-based framework to obtain training and validation datasets from seafloor DAS measurements. We demonstrate that this method is useful to identify earthquake signals using a simple template, and to identify P and S arrival times from scratch using common seismic picking methods. The novelty of this method relies in the recovery of P and S pickings from most of the DAS channels using SVR. The uncertainty of the pickings is based on the SNR of each channel.\u003c/p\u003e\u003cp\u003eThe availability of DAS-based databases is vital for developing deep learning models for seafloor DAS data. Self-noise in DAS data is due to signal fading, laser noise, and common-mode noise (Lindsey et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Farghal et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Additionally, seafloor DAS measurements present additional sources of noise. Common 2D filtering techniques are useful to identify and extract earthquake signals from strain rate measurements. This has been done to identify surface waves to obtain P and S velocity models (Spica et al., \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Fukushima et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). However, nonphysical signals may have lie in the same frequency band as the extracted signal.\u003c/p\u003e\u003cp\u003eThe source of this noise may lie in the cable coupling into the seafloor. Most of the measurements are done using telecommunication fibers deployed for signal transmission. Hence, the coupling may not be the same throughout the cable (Farghal et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). It has been demonstrated that the strain amplitude can change significantly with the degree of coupling of the cable (Ajo-Frankin et al., 2019; Lindsey and Martin, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). For this reason, DAS-based datasets for deep learning must address these types of noise.\u003c/p\u003e\u003cp\u003eAlthough our method has been proven useful, there are some limitations. First, the kernel choice for SVR is based on the number of training points distributed throughout the DAS channels, and the interquartile range of the difference between the arrival times of P and S. Although the examples shown demonstrated that this criterion works, there may be events that do not fulfill these criteria. Current work is based on analyzing more DAS data to establish a more robust criterion for the kernel selection.\u003c/p\u003e\u003cp\u003eDeeping more in the kernel\u0026rsquo;s discussion, so far, we have Radial Basis Functions (RBF) and linear kernels for obtaining picking models. The main reason for using these two are based on the following points: 1) Time calculations, and 2) distribution of pickings across the DAS data. Earthquake events whose source is close to the cable will produce nonlinear arrival times distribution, whereas farther events will display a linear distribution with different slopes due to the different apparent velocities of P and S. The time execution of SVR algorithms depends on the number of points, but it ranges between 0.3\u0026ndash;5.0 s. Linear kernels are more time expensive than RBF kernels. It is possible that other kernels such as polynomial, sigmoid, etc. may produce a better adjustment. However, this may result in an increase in the time calculation.\u003c/p\u003e\u003cp\u003eAnother limitation lies in the picking methods. Overall, STA/LTA and AR-AIC methods have good agreement in relatively high SNR signals and disagrees in when decreasing the SNR. The reason for the agreement is that prior to the estimation of the autoregressive models, the method applies STA/LTA for P and S phrases (Akazawa et al., 2004). However, we found events such as 2022-02-28 16:52:20 UTC that failed to predict. Although we observe in the DAS section (see Figure S5 \u0026ndash; S6 from Supplemental Material) two clear arrivals, the pickings obtained from both methods completely mismatch. Further research will be done to obtain a more robust implementation of this framework.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eHere we present SVR-DAS, a ML-based method to create training and validation datasets from seafloor Distributed Acoustic Sensing measurements. Our method is divided into two stages: earthquake detection and seismic phase picking. In the first stage, our method applies envelope template matching to identify earthquake strain rate measurements in DAS channels. The template matching can start from one template, and the algorithm feedback the template library with the new detections. In the second stage, we apply STA/LTA (Allen, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1982\u003c/span\u003e) and autoregressive models using the Akaike Information Criterion (AR-AIC) (Akazawa et al., 2014) to pick P and S arrival times from the DAS channels. These pickings are compared to each other, and the coherent pickings are used as training sets to obtain P and S pick models using Support Vector Regression (SVR) (Vapnik, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e1997\u003c/span\u003e). Because P-to-S converted waves have higher amplitude than the true P wave, most of the pickings are associated with the converted wave. For this reason, we apply a spectrogram-based picker using a time window near the P-to-S converted waves in order to identify the time where the signal power increases due to the P wave arrival. We tested our method using relatively high SNR DAS data associated with an earthquake that occurred in 2023-07-23 successfully retrieving both P and S arrival times. Afterwards, we applied SVR-DAS using the data from 2022-02-28 08:51:20 h to 23:59:20 h UTC. Our method identified 9 events with difference in P and S time arrivals less than 20 s. From these events, we retrieved approximately around 450 000 DAS channels with identified P and S time arrivals. Using a subset of 18 000 examples (12 000 events and 6000 noise), we train a CNN-RNN model to identify P and S seismic pickings, achieving an accuracy of 96%. We believe that the generated DAS-based databases obtained from SVR-DAS will be useful to develop state-of-the-art deep learning models, or fine tuning already existent deep learning models, to explore the seafloor DAS measurements. A potential application of this method is also in Earthquake Early Warning system, that relies on accurate P wave estimation to determine ground motion shaking.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAcknowledgments\u003c/h2\u003e\u003cp\u003eThis study is supported by the MEXT project for Seismology Toward Research Innovation with Data of Earthquakes (STAR-E) No. JPJ010217. The key ideas in this study were derived from the activities of the MEXT Volcano Practical Human Resource Development Support Program Japan, the MEXT The Third Earthquake and Volcano Hazards Observation and Research Program (Earthquake and Volcano Hazard Reduction Research), JST NEXUS, Japan No. JPMJNX25C5, the JSPS KAKENHI Grant-in-Aid for Scientific Research (A) No. 23H00466, Fund for the Promotion of Joint International Research (International Collaborative Research) No. 23KK0181, Grant-in-Aid for Challenging Research (Pioneering) No. 25K21806, and Earthquake Research Institute, The University of Tokyo Joint Research ERI JURP 2025-A-03, 2024-B-01, and 2025-B-01. We thank Dr. Shinichi Ito, Dr. Tomoki Tokuda, and Hiroaki Yamahana for their valuable feedback regarding this study. We thank T. Yagi from ERI, The University of Tokyo for lending us the data on Sanriku cable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAllen R (1982) Automatic phase pickers: Their present use and future prospects. Bull. Seis. Soc. Am. 72(6B): S225 \u0026ndash; S242\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAjo-Franklin JB, Dou S, Lindsey NJ, Monga I, Tracey C, Robertson M (2019) Distributed acoustic sensing using dark fiber for near surface characterization and broadband seismic event detection. Sci Rep 9(1):1328. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-018-36675-8\u003c/span\u003e\u003cspan address=\"10.1038/s41598-018-36675-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAkazawa T (2004) A technique for automatic detection of onset time of P- and S- Phases in strong motion records\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAoi S, Asano Y, Kunugi T, Kimura T, Uehira K, Takahashi N, Ueda H, Shiomi K, Matsumoto T, Fujiwara T (2020) MOWLAS: NIED observation network for earthquake, tsunami and volcano. Earth Planet Space 72(126). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s40623-020-01250-x\u003c/span\u003e\u003cspan address=\"10.1186/s40623-020-01250-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAoi S, Takeda T, Kunugi T, Shinohara M, Miyoshi T, Uehira K (2023) Development and Construction of Nankai Trough Seafloor Observation Network for Earthquakes and Tsunamis: N-Net. 2023 IEEE Underwater Techonology. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/UT49739.2023.10103206\u003c/span\u003e\u003cspan address=\"10.1109/UT49739.2023.10103206\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAtterholt J, Zhan Z, Shen Z, Li Z (2022) A unified wavefield-partitioning approach for distributed acoustic sensing. Geophys J Int 228(2):1410\u0026ndash;1418. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/gji/ggab-407\u003c/span\u003e\u003cspan address=\"10.1093/gji/ggab-407\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBaba S, Araki E, Yokobiki T, Kawamata K, Uchiyama K, Yoshizuka T (2024) Seismic observation using distributed acoustic sensing around the Tsugaru Strait at the Japan and Kuril Trenches, northeastern Japan. Earth Planet Space 76(28). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s40623-024-01975-z\u003c/span\u003e\u003cspan address=\"10.1186/s40623-024-01975-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBenioff H (1935) A linear strain seismograph. Bull Seism Soc Am 25(4):283\u0026ndash;309\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCastongia E, Wang HF, Lord N, Fratta D, Mondanos M, Chalari A (2017) An experimental investigation of distributed acoustic sensing (DAS) on lake ice. J Environ Eng Geophys 22:167\u0026ndash;176. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2113/JEEG22.2.167\u003c/span\u003e\u003cspan address=\"10.2113/JEEG22.2.167\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDaley TM, Freifeld BM, Ajo-Franklin J, Dou S, Pevzner R (2013) Field testing of fiber-optic distributed acoustic sensing (DAS) for subsurface seismic monitoring. Lead Edge 32(6):593\u0026ndash;724. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1190/tle32060699.1\u003c/span\u003e\u003cspan address=\"10.1190/tle32060699.1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDing Q, Shen Z, Zhu W, Liu B (2025) DASFormer: self-supervised pretraining for earthquake monitoring. Visual Intell 3(14). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s44267-025-00085-y\u003c/span\u003e\u003cspan address=\"10.1007/s44267-025-00085-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFarghal NS, Saunders JK, Parker GA (2022) The Potential of Using Fiber Optic Distributed Acoustic Sensing (DAS) in Earthquake Early Warning Applications. Bull Seism Soc Am 115(2):516\u0026ndash;532. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1785/0120240234\u003c/span\u003e\u003cspan address=\"10.1785/0120240234\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFujie G, Kodaira S, Kaiho Y, Yamamoto Y, Takahashi T, Miura S, Yamada T (2018) Controlling Factor of incoming plate hydration at the north-western Pacific margin. Nat Commun 9(3844). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-018-06320-z\u003c/span\u003e\u003cspan address=\"10.1038/s41467-018-06320-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFukushima S, Shinohara M, Nishida K, Takeo A, Yamada T, Yomogida K (2022) Detailed S-wave velocity structure of sediment and crust off Sanriku, Japan by a new analysis method for distributed acoustic sensing data using a seafloor cable and seismic interferometry. Earth Planet Space 74(92). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s40623-022-01652-z\u003c/span\u003e\u003cspan address=\"10.1186/s40623-022-01652-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFukushima S, Shinohara M, Yamada T, Hino R, Azuma R, Ito Y, Yamashita Y, Takano H (2025) Enhanced P-wave velocity imaging by marine airgun-source seismic surveys with distributed acoustic sensing. Sci Rep 15(18111). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-025-01190-0\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-01190-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGibbons SJ, Ringdal F (2006) The detection of low magnitude seismic events using array-based waveform correlation. Geophys J Int 165(1):149\u0026ndash;166. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/j.1365-246X.2006.02865.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1365-246X.2006.02865.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHartog AH (2017) An Introduction to Distributed Optical Fibre Sensors. CRC, Boca Raton\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHough SE (1990) Constraining sediment thickness in the San Francisco Bay Area Using Observed Resonances and P-to-S Conversions. Geophys Res Lett 17(9):1469\u0026ndash;1472. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1029/GL017i009p01469\u003c/span\u003e\u003cspan address=\"10.1029/GL017i009p01469\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKaneko R, Nagao H, Ito S, Tsuruoka H, Obara K (2023) Detection of Deep Low-Frequency Tremors From Continuous Paper Records at a Station in Southwest Japan About 50 Years Ago Based on Convolutional Neural Network. J Geophys Res Sol Earth 128(2). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1029/2022JB024842\u003c/span\u003e\u003cspan address=\"10.1029/2022JB024842\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKanazawa T, Uehira K, Mochizuki M, Shinbo T, Fujimoto H, Noguchi S, Kunugi T, Shiomi K, Aoi S, Matsumoto T, Sekiguchi S, Okada T (2016) S-net project, cabled observation network for earthquakes and tsunamis. SubOptics 2016, WE2B-3\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKatoh S, Iio Y, Nagao H, Katao H, Sawada M, Tomikasa K (2025) SegPhase: development of arrival time picking models for Japan\u0026rsquo;s seismic network using the hierarchical vision transformer. Earth Planet Space 77(118). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s40623-025-02249-y\u003c/span\u003e\u003cspan address=\"10.1186/s40623-025-02249-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKatoh S, Nagao H, Iio Y (2025) PoViT-UQ: P-wave Polarity and Arrival Time Determination using Vision Transformer with Uncertainty Quantification. Geophys J Int. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/gji/ggaf324\u003c/span\u003e\u003cspan address=\"10.1093/gji/ggaf324\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKawaguchi K, Araki E, Kaneda Y (2011) Establishment for a method for real-time and long-term seafloor monitoring. J Adv Mar Sci Technol Soc 17:125\u0026ndash;135. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.14928/amstec.17.2_125\u003c/span\u003e\u003cspan address=\"10.14928/amstec.17.2_125\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eK\u0026uuml;perkoch L, Meier T, Diehl T (2011) Automated Event and Phase Identification. In: Bormann, P. (Ed.), New Manual of Seismological Observatory Practice 2 (NMSOP-2), Postdam: Deutsches GeoForschungsZentrum GFZ, 1\u0026ndash;52. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2312/GFZ.NMSOP-2_ch16\u003c/span\u003e\u003cspan address=\"10.2312/GFZ.NMSOP-2_ch16\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi Z, Zhan Z (2018) Pushing the limit of earthquake detection with distributed acoustic sensing and template matching: a case study at the Brady geothermal field. Geophys J Int 215(3):1583\u0026ndash;1593. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/gji/ggy359\u003c/span\u003e\u003cspan address=\"10.1093/gji/ggy359\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLindsey NJ, Martin ER, Dreger DS, Freifeld B, Cole S, James SR, Biondi BL, Ajo-Franklin JB (2017) Fiber-Optic Network Observations of Earthquake Wavefields. Geophys Res Lett 44(23):11792\u0026ndash;11799. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/2017GL075722\u003c/span\u003e\u003cspan address=\"10.1002/2017GL075722\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLindsey NJ, Martin RM (2020) Fiber-optic seismology. Annu Rev Earth Planet Sci 49:309\u0026ndash;336. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1146/annurev-earth-072420-065213\u003c/span\u003e\u003cspan address=\"10.1146/annurev-earth-072420-065213\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLindsey NJ, Rademacher H, Ajo F (2020) On the broadband instrument response of fiber-optic DAS arrays. J. Geophys. Res. 125(e2019JB018145). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1029/2019JB018145\u003c/span\u003e\u003cspan address=\"10.1029/2019JB018145\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu B, Yang S, Ren Y, Xu X, Jiang P, Chen Y (2021) Deep-learning seismic full-waveform inversion for realistic structural models. Geophysics 86(1). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1190/geo2019-0435.1\u003c/span\u003e\u003cspan address=\"10.1190/geo2019-0435.1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMatsubara M, Obara K, Kasahara K (2008) Three-dimensional P- and S-wave velocity structures beneath the Japan Islands obtained by high-density seismic stations by seismic tomography. Tectonophysics 454(1\u0026ndash;4):86\u0026ndash;103. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.tecto.2008.04.016\u003c/span\u003e\u003cspan address=\"10.1016/j.tecto.2008.04.016\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMasoudi A, Newson TP (2016) Contributed review: distributed optical fibre dynamic strain sensing. Rev Sci Instrum 87:011501. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1063/1.4939482\u003c/span\u003e\u003cspan address=\"10.1063/1.4939482\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMcBrearty I, Beroza GC (2023) Earthquake Phase Association with Graph Neural Networks. Bull Seism Soc Am 113(2):524\u0026ndash;547. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1785/0120220182\u003c/span\u003e\u003cspan address=\"10.1785/0120220182\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMiao Y, Salaree A, Spica ZJ, Nishida K, Yamada T, Shinohara M (2025) Assessing the Earthquake Recording Capability of an Ocean-Bottom Distributed Acoustic Sensing Array in the Sanriku region, Japan. Seis Res Lett 96:631\u0026ndash;650. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1785/0220240120\u003c/span\u003e\u003cspan address=\"10.1785/0220240120\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMendo-P\u0026eacute;rez G, Nagao H, Katoh S, Shinohara M (2025) Leveraging CNN-RNN-based model for earthquake detection in Distributed Acoustic Sensing measurements [Work in preparation]. The University of Tokyo\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMousavi SM, Ellsworth WL, Zhu W, Chuang LY, Beroza GC (2020) Earthquake transformer \u0026ndash; an attentive deep-learning model for simultaneous earthquake detection and phase picking. Nat Commun 11(3952). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-020-17591-w\u003c/span\u003e\u003cspan address=\"10.1038/s41467-020-17591-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMoseley B, Nissen-Meyer T, Markham A (2020) Deep learning for fast simulation of seismic waves in complex media. Solid Earth 11(4):1527\u0026ndash;1549. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5194/se-11-1527-2020\u003c/span\u003e\u003cspan address=\"10.5194/se-11-1527-2020\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMuir JB, Zhan Z (2022) Wavefield-based evaluation of DAS instrument response and array design. Geophys J Int 229(1):21\u0026ndash;34. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/gji/ggab439\u003c/span\u003e\u003cspan address=\"10.1093/gji/ggab439\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRoss ZE, Meier MA, Hauksson E, Heaton TH (2018) Generalized Seismic Phase Detection with Deep Learning. Bull Seismo Soc Am 108(5A):2894\u0026ndash;2901. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1785/0120180080\u003c/span\u003e\u003cspan address=\"10.1785/0120180080\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRoss ZE, Yue Y, Meier MA, Hauksson E, Heaton TH (2019) PhaseLink: A Deep Learning Approach to Seismic Phase Association. J Geophys Res Sol Ea 124(1):856\u0026ndash;869. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1029/2018JB016674\u003c/span\u003e\u003cspan address=\"10.1029/2018JB016674\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShinohara M, Yamada T, Akuhara T, Mochizuki K, Sakai S, Hamakawa M, Kasajima T, Arioka T, Kubota S (2019) Distributed Acoustic Sensing measurement by using seafloor optical fiber cable system off Sanriku for seismic observation. Paper presented at OCEANS 2019 MTS/IEEE SEATTLE. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.23919/OCEANS40490.2019.8962757\u003c/span\u003e\u003cspan address=\"10.23919/OCEANS40490.2019.8962757\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShinohara M, Yamada T, Akuhara T, Mochizuki K, Sakai S (2022) Performance of a Seismic Observation by Distributed Acoustic Sensing Technology Using a Seafloor Cable Off Sanriku, Japan. Front Mar Sci 9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fmars.2022.844506\u003c/span\u003e\u003cspan address=\"10.3389/fmars.2022.844506\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSpica ZJ, Nishida K, Akuhara T, Petrelis F, Shinohara M, Yamada T (2020) Marine Sediment Characterized by Ocean-Bottom Fiber-Optic Seismology. Geophys Res Lett 47(16). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1029/2020GL088360\u003c/span\u003e\u003cspan address=\"10.1029/2020GL088360\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSpica ZJ, Castellanos JC, Viens L, Nishida K, Akuhara T, Shinohara M, Yamada T (2022) Subsurface Imaging With Ocean-Bottom Distributed Acoustic Sensing and Water Phases Reverberations. Geophys Res Lett 49(2). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1029/2021GL095287\u003c/span\u003e\u003cspan address=\"10.1029/2021GL095287\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSupriya K, Divya A, Vinodkumar B, Ram Sai G (2020) Trigger Word Recognition using LSTM. Int J Eng Res Technol 9(6)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTakahashi N, Kodaira S, Tsuru T, Park J, Kaneda Y, Suyehiro K, Kinoshita H, Abe S, Nishino M, Hino R (2004) Seismic structure and seismogenesis off Sanriku region, northeastern Japan. Geophys J Int 159(1):129\u0026ndash;145. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/j.1365-246X.2004.02350.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1365-246X.2004.02350.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTokuda T, Nagao H (2023) Seismic-phase detection using multiple deep learning models for global and local representations of waveforms. Geophys J Int 235(2):1163\u0026ndash;1182. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/gji/ggad270\u003c/span\u003e\u003cspan address=\"10.1093/gji/ggad270\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVapnik VN (1997) The support vector method. In International conference on artificial neural networks (pp. 261\u0026ndash;271). Berlin, Heidelberg. Springer Berlin Heidelberg\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eViens L, Bonilla LF, Spica ZJ, Nishida K, Yamada T, Shinohara M (2022) Nonlinear Earthquake Response of Marine Sediments With Distributed Acoustic Sensing. Geophys Res Lett 49(21). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1029/2022GL100122\u003c/span\u003e\u003cspan address=\"10.1029/2022GL100122\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWalter F, Graff D, Lindner F, Paitz P, Koppfli M, Chmiel M, Fichtner A (2020) Distributed acoustic sensing of microseismic sources and wave propagation in glaciated terrain. Nat Commun 11:2436. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-020-15824-6\u003c/span\u003e\u003cspan address=\"10.1038/s41467-020-15824-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXiao H, Tilmann F (2024) DeepDAS: An Earthquake Phase Identification Tool Using Submarine Distributed Acoustic Sensing Data, Galileo conference: Fibre Optic Sensing in Geosciences, Catania, Italy, 16\u0026ndash;20 Jun 2024, GC12-FibreOptic-51. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5194/egusphere-gc12-fibreoptic-51\u003c/span\u003e\u003cspan address=\"10.5194/egusphere-gc12-fibreoptic-51\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYin J, Zhu W, Li J, Biondi E, Miao Y, Spica ZJ, Viens L, Shinohara M, Ide S, Mochizuki K, Husker AL, Zhan Z (2023) Earthquake Magnitude With DAS: A Transferable Data-Based Scaling Relation. Geo Phys Lett 50(10). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1029/2023GL103045\u003c/span\u003e\u003cspan address=\"10.1029/2023GL103045\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYu S, Ma J, Wang W (2019) Deep learning for denoising. Geophysics 84(6). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1190/geo2018-0668.1\u003c/span\u003e\u003cspan address=\"10.1190/geo2018-0668.1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYu C, Zhan Z, Lindsey NJ, Ajo-Franklin JB, Robertson M (2019) The Potential of DAS in Teleseismic Studies: Insights From the Goldstone Experiment. Geophys Res Lett 46(3):1320\u0026ndash;1328. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1029/2018GL081195\u003c/span\u003e\u003cspan address=\"10.1029/2018GL081195\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhan Z (2019) Distributed Acoustic Sensing. Turns Fiber-Optic Cables into Sensitive Seismic Antennas. Seism Res Lett 91(1):1\u0026ndash;15. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1785/0220190112\u003c/span\u003e\u003cspan address=\"10.1785/0220190112\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhu W, Beroza GC (2019) PhaseNet: a deep-neural-network-based seismic arrival-time picking method. Geophys J Int 216(1):261\u0026ndash;273. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/gji/ggy42\u003c/span\u003e\u003cspan address=\"10.1093/gji/ggy42\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhu W, Biondi E, Li J, Ross ZE, Zhan Z (2023) Seismic arrival-time picking on distributed acoustic sensing data using semi-supervised learning. Nat Commun 14(8192). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41467-023-4335-3\u003c/span\u003e\u003cspan address=\"10.1038/s41467-023-4335-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhu W, McBrearty I, Mousavi SM, Ellsworth WL, Beroza GC (2022) Earthquake Phase Association Using a Bayesian Gaussian Mixture Model. J Geophys Res Sol Earth 127(5). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1029/2021JB023249\u003c/span\u003e\u003cspan address=\"10.1029/2021JB023249\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang J, Li Z, Zhang J (2023) Simultaneous Seismic Phase Picking and Polarity Determination with an Attention-Based Neural Network. Seism Res Lett 94(2A):813\u0026ndash;828. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1785/0220\u003c/span\u003e\u003cspan address=\"10.1785/0220\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[{"identity":"f32a59f1-6dfd-4317-8890-c0064004d991","identifier":"10.13039/501100001700","name":"Ministry of Education, Culture, Sports, Science and Technology","awardNumber":"JPJ010217","order_by":0},{"identity":"a83a24c2-5cae-47a3-913d-17c3428d78be","identifier":"10.13039/501100001700","name":"Ministry of Education, Culture, Sports, Science and Technology","awardNumber":"JPMJNX25C5","order_by":1},{"identity":"bb949a0a-fa4c-44a1-9a82-292819d14a46","identifier":"10.13039/501100001691","name":"Japan Society for the Promotion of Science","awardNumber":"23H00466","order_by":2},{"identity":"f4f14a76-0289-451a-8f89-75ef2e5a38b0","identifier":"10.13039/501100001691","name":"Japan Society for the Promotion of Science","awardNumber":"23KK0181","order_by":3},{"identity":"ab4acfa9-cb8e-4054-8a87-92f9c37a8900","identifier":"10.13039/501100001691","name":"Japan Society for the Promotion of Science","awardNumber":"25K21806","order_by":4},{"identity":"5b6aede8-4ffd-4059-90d1-851cd8abc7bc","identifier":"10.13039/501100014773","name":"Earthquake Research Institute, University of Tokyo","awardNumber":"ERI JURP 2025-A-03","order_by":5},{"identity":"1a0db96e-7f1f-40f9-b000-1fd73b594ae5","identifier":"10.13039/501100014773","name":"Earthquake Research Institute, University of Tokyo","awardNumber":"ERI JURP 2024-B-01","order_by":6},{"identity":"2e4e0671-83dc-49e5-8eed-8ef3a67197dc","identifier":"10.13039/501100014773","name":"Earthquake Research Institute, University of Tokyo","awardNumber":"ERI JURP 2025-B-01","order_by":7}],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Earthquake Research Institute, The University of Tokyo","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Distributed Acoustic Sensing, Deep Learning, Support Vector Regression, earthquake detection, seismic picking","lastPublishedDoi":"10.21203/rs.3.rs-8080988/v2","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8080988/v2","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eSeafloor earthquake monitoring is one of the emergent research areas in seismology. Its importance relies on the continuous monitoring of earthquake activity near or on subduction zones, where large megathrust earthquakes are generated. In Japan, seafloor observation systems have been deployed to monitor earthquake activity in Nankai area (DONET) and Japan Trench (S-net). However, these are limited to the number of observation points of the system, and their installation and maintenance can be logistically and economically expensive. Distributed Acoustic Sensing (DAS) measurements obtained from fiber optic cables are arising as a promising technology to monitor offshore earthquake activity. In addition, DAS data is suitable for deep learning methods due to the vast volume of data that it produces. Although successful methods have been developed for land-based DAS measurements, their performance in seafloor DAS data is compromised by additional sources of noise and coupling issues. State-of-the-art research is aimed at developing novel deep-learning-based methods to process seafloor data. However, DAS-based earthquake catalogs are still scarce, and manual picking is not feasible. In this work, we propose a novel machine-learning-based method to build earthquake catalogs from seafloor DAS measurements. Our method detects earthquakes by applying envelope template matching using a small number of templates obtained from DAS data itself. P and S arrival times are obtained from detections using STA/LTA and Autoregressive models using the Akaike Information Criterion (AR-AIC). The coincident picks of both methods are used as training data to create P and S picks models using Support Vector Regression (SVR). For this reason, our method is called SVR-DAS. We test the method using relatively high Signal-to-Noise Ratio (SNR) DAS data from an earthquake that occurred in 2023-07-26. Subsequently, we applied SVR-DAS to the DAS data from 2022-02-28, obtaining P and S models from relatively high SNR DAS data. The created catalogs are used to train and validate a CNN-RNN model. In addition, we present a comparison of the SVR-DAS picks with the picks obtained from PhaseNet-DAS, a deep-learning-based model for earthquake detection and seismic picking. Our results prove that SVR-DAS is capable to build earthquake catalogs from DAS data with enough accuracy. The obtained catalogs are useful for further development of deep learning models for seafloor DAS data processing.\u003c/p\u003e","manuscriptTitle":"SVR-DAS: A Machine-Learning-Based Method to Create Earthquake Catalogs from Seafloor Distributed Acoustic Sensing Measurements","msid":"","msnumber":"","nonDraftVersions":[{"code":2,"date":"2025-11-24 18:04:11","doi":"10.21203/rs.3.rs-8080988/v2","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}},{"code":1,"date":"2025-11-12 04:25:36","doi":"10.21203/rs.3.rs-8080988/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"10205052-c2c1-4173-b175-5f83c9bb726f","owner":[],"postedDate":"November 24th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":57838552,"name":"Seismology"},{"id":57838553,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2026-03-02T23:29:07+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-24 18:04:11","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v2","identity":"rs-8080988","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8080988","identity":"rs-8080988","version":["v2"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}