Enhanced deep learning approach for detecting and locating tectonic tremors in the Nankai subduction zone

Enhanced deep learning approach for detecting and locating tectonic tremors in the Nankai subduction zone | 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 Enhanced deep learning approach for detecting and locating tectonic tremors in the Nankai subduction zone Yuya Jinde, Amane Sugii, Yoshihiro Hiramatsu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6526402/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 22 Jul, 2025 Read the published version in Earth, Planets and Space → Version 1 posted 5 You are reading this latest preprint version Abstract Tectonic tremors are key indicators of slow-slip phenomena, and detecting them accurately is a challenging task. Conventional techniques often fail to detect tremors during periods of intense tremor activity. We present here a deep learning approach for detecting and locating tremors in the Nankai subduction zone that is more effective than conventional techniques. We utilized two convolutional neural networks (CNNs): a CNN for classification of seismic waveforms into noise, tremors, or earthquakes, and a CNN for regression prediction of tremor epicenters from amplitude data. The accuracy, recall, and precision of the CNN for classification all exceeded 95%. The CNN for regression used ensemble predictions to produce estimates of tremor locations with a median error of 3.2 km. When this approach was applied to continuous data, it successfully mapped key features of tremor activity and improved the detection and location of tremors, especially during high-activity periods. convolutional neural network slow slip tremor migration rapid tremor reversal Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction Slow earthquakes, characterized by lower fault slips than regular earthquakes, occur mainly along plate boundaries such as the Nankai subduction zone. They are classified into geodetically detected slow-slip events (SSEs), seismically observed low-frequency earthquakes, very low-frequency earthquakes, and tectonic tremors. Short-term SSEs last weeks to months and recur every few months. Their recurrence intervals and migration patterns may change before megathrust events (e.g., Matsuzawa et al. 2010 ). These events may be linked to megathrust earthquakes because they occur near megathrust rupture zones (e.g., Obara and Kato 2016 ). Interactions have been suggested to occur between slow slips and megathrust earthquakes in subduction zones (Kato et al. 2012 ; Ruiz et al. 2014 ). Tectonic tremors, weak vibrations of 1–10 Hz lasting minutes to hours, often coincide with short-term SSEs to form Episodic Tremors and Slip (e.g., Rogers and Dragert 2003 ). Because tremors serve as proxies for short-term SSEs and slip rates at plate interfaces (e.g., Hiramatsu et al. 2008 ; Obara 2011 ), precise and continuous monitoring is crucial for understanding the link between slow-slip and megathrust events. However, it is challenging to detect and locate tremors because of the lack of clear P- and S-wave arrivals. The envelope correlation method (ECM) (Obara 2002 ), which is based on waveform similarity, is commonly used for detecting and locating tremors and has been refined. Maeda and Obara ( 2009 ) improved location accuracy by incorporating radiation energy, and Mizuno and Ide ( 2019 ) redefined the ECM using a maximum likelihood method that enhanced the accuracy and enabled the detection of multiple sources. However, ECM-based methods often misclassify regular earthquakes and noise as tremors. High tremor activities or noise levels cause a reduction of waveform similarity that leads to missed detections. Advances in machine learning, particularly deep learning, have led to novel findings in seismology (Kong et al. 2019 ; Mousavi and Beroza 2022 ; Kubo et al. 2024 ). Convolutional neural networks (CNNs) have been widely used in earthquake analysis. Kriegerowski et al. ( 2019 ) have developed a CNN-based regression model for locating hypocenters of earthquakes from seismic waveforms without phase detection. Thomas et al. ( 2021 ) have constructed a CNN for detecting low-frequency earthquakes on the Parkfield section of the San Andreas fault. CNNs have also been used to classify tremors and other events in the Nankai Trough (Nakano et al. 2019 ), the Japan Trench (Takahashi et al. 2021 ), and the Cascadia subduction zone (Rouet-Leduc et al. 2020 ). However, these CNN-based approaches face challenges of misclassifications between tremors, distant earthquakes, and noise. Recently, Sugii et al. ( 2024 ) have included distant earthquakes in training data and have developed a highly accurate CNN-based approach to classify waveforms recorded by a temporal array network in the Nankai subduction zone, although their analysis has been limited to tremors around the Kii Peninsula. A new method for detecting and locating tremors using permanent stations, independent of waveform similarity and applicable across the entire Nankai subduction zone, is therefore needed. In this study, we developed a novel CNN-based approach for detecting and locating tremors in the Nankai subduction zone, consisting of a CNN for classifying events (CNN-classification) and a CNN for regression prediction of tremor locations (CNN-regression). This method outperformed conventional techniques and successfully captured previously reported spatiotemporal features of tremors. 2. Data and Methods We used three-component velocity waveforms recorded at 129 Hi-net stations operated by the National Research Institute for Earth Science and Disaster Resilience around the Nankai subduction zone from January 2008 to September 2016 (Fig. 1 a). 2.1 Classification of events 2.1.1 Datasets for the CNN-classification We constructed a dataset with three categories: noise, tremors, and earthquakes. Tremor data were identified by reference to the origin times and epicenters in the tremor catalog of Mizuno and Ide ( 2019 ). Periods with at least five tremors per hour were identified, and waveforms were visually inspected. Only confirmed tremor waveforms were included. Waveforms from stations over 40 km from the epicenter were excluded to maintain classification accuracy (Sugii et al. 2024 ). For unlisted times, epicentral distances were estimated based on the average locations of tremors that occurred within one hour. Earthquake data, such as hypocenters and arrival times, were obtained from catalogs reported by the Japan Meteorological Agency. To preserve classification accuracy, events recorded at any of the 129 stations were included, except for those shallower than 50 km with magnitudes less than 1 or deeper than 50 km with magnitudes less than 3. The start times of one-minute windows were randomly shifted by up to 45 seconds preceding the arrival times of the P-wave to enhance model robustness, and one-minute waveforms were extracted. Waveforms that were not classified as tremors or earthquakes were treated to be noise. For noise data selection, visual inspections at multiple stations ensured the exclusion of waveforms with abnormal signals or features indicating tremors or earthquakes. To balance the data, we applied random sampling and excluded waveforms with missing components. Each category (noise, tremors, and earthquakes) consequently obtained 126571 spectrograms. Figure 1 b and 1 d show the distributions of the epicenters of tremors and earthquakes used for the CNN-classification. Training and validation data spanned the period from January 2008 to December 2015 and consisted of 104091 spectrograms per category. Ninety percent of these spectrograms were randomly chosen for training, and the remainder were used for validation. The test data consisted of 22480 spectrograms per category and covered the period from January to September 2016 (Figs. S1 and S2). We prepared input data for the CNN-classification following Sugii et al. ( 2024 ). Spectrograms were generated from 2–8 Hz bandpass-filtered, three-component waveforms of one-minute duration, sampled at 5-second intervals with a lag of 1.19 seconds (Fig. 2 a and 2 b). Each spectrogram consisted of 52 × 52 pixels and covered 0–10 Hz on a linear scale. Amplitude spectra were normalized by subtracting the mean and dividing by the standard deviation. 2.1.2 Architecture and training of CNN-classification The CNN-classification followed the architecture of Sugii et al. ( 2024 ), which includes three convolutional layers with 3 × 3 filters, two MaxPooling layers, two dropout layers, and one softmax layer, along with fully connected layers of 32 and 3 nodes (Fig. S3a). Rectified Linear Units were used as the activation function. The batch size and learning rate were set to 64 and 10 − 4 , respectively, based on the most accurate model of Sugii et al. ( 2024 ). We used the Adam algorithm (Kingma and Ba 2015 ) to apply the categorical cross-entropy loss function. Training stopped if the validation loss did not improve during four consecutive epochs or after 500 epochs. The model with the lowest validation loss was used for testing. For each inputted one-minute spectrogram, the CNN output the probabilities of classifying the waveform as noise, tremor, or earthquake. 2.2 Locating tremor epicenters The spatial distribution of amplitudes facilitated locating tremors (e.g., Maeda and Obara 2009 ). Furthermore, the tremor probabilities output by the CNN-classification provided information that facilitated estimating tremor epicenters because they tended to be high and form clusters near the epicenter (Fig. 2 c). In this study, we used tremor probabilities for waveform preprocessing and developed the CNN-regression that determined tremor epicenters based on amplitude data. 2.2.1 Dataset for CNN-regression We applied the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm (Ester et al. 1996 ) to detect clusters of stations with high tremor probabilities based on the CNN-classification. For cataloged times, clustering using DBSCAN was performed on stations with tremor probabilities of 0.9 or higher, using ε = 0.5° and MinPts = 3, where ε is the radius of the neighborhood, and MinPts is the minimum number of stations. These parameters were determined through trial and error. We selected events with clusters for training, validation, and testing. If multiple clusters were formed simultaneously, separate data were created for each cluster. We excluded events with fewer than three stations, at which tremor probabilities were 0.9 or greater, within 0.5° of the epicenters. In addition, we excluded data from 2012 to conduct the continuous data analysis. The final dataset consisted of 116136 events (Fig. 1 c) divided into training (76523: January 2008 to December 2013 except for 2012), validation (15573: January 2014 to December 2014), and test (24040: January 2015 to September 2016) data (Fig. S4). The CNN-regression used RMS amplitudes from bandpass-filtered (2–8 Hz), one-minute waveforms. Epicenters were taken from the tremor catalog of Mizuno and Ide ( 2019 ) and served as ground truth. The input data consisted of a 3 × 129 matrix that represented the root mean square (RMS) amplitude of the north–south, east–west, and vertical components of the waveforms at 129 stations. To reduce the impact of abnormal amplitudes, values exceeding three standard deviations above the mean RMS amplitude were corrected to 10 − 9 m/s, comparable to the average noise level. Similarly, amplitude corrections were applied to stations with tremor probabilities less than 0.9, to those classified as non-clusters by DBSCAN, and to those assigned to non-target clusters. Finally, the amplitudes were standardized for each component by subtracting the mean and dividing by the standard deviation to ensure uniform scaling. 2.2.2 Architecture and training of CNN-regression The CNN-regression consisted of two convolutional layers with 3 × 2 filters, followed by fully connected layers with 128, 64, and 2 nodes (Fig. S3b). The omission of pooling layers to preserve spatial information enabled the model to implicitly learn the locations of stations. Epicenters from the catalog were converted into latitude and longitude differences relative to an origin at 34.1°N, 135°E for stable and efficient training. The output layer predicted the latitude and longitude relative to this origin. The CNN-regression was trained using the mean absolute error as the loss function and the Adam algorithm with a batch size of 32 and a learning rate of 5 × 10 − 4 . Training stopped if the loss did not improve for 30 consecutive epochs or after 400 epochs. We used ensemble learning with 100 models to assess prediction uncertainty. Each model with randomly initialized parameters was trained to minimize validation losses. The average and standard deviation of those predictions were treated as the final model prediction and uncertainty, respectively. 2.2.3 Application to continuous data For continuous data, spectrograms at one-minute intervals for 129 stations were input into the CNN-classification. If DBSCAN formed clusters of stations with tremor probabilities of 0.9 or greater, the data with their RMS amplitude corrections were input into the CNN-regression to locate epicenters. When multiple clusters were detected, the epicenter was estimated for each cluster. Events exceeding the uncertainty threshold determined from the test data (see section 3.2 for details) were excluded as outliers. In addition, events were excluded if they were detected at fewer than three stations, at which tremor probabilities were 0.9 or greater, located within a 0.5° square of the estimated epicenter. Events were also excluded as isolated if no adjacent events were detected within 5 km during a 12-hour window before or after the event (Nakamoto et al. 2021 ). 3. Results and Discussion 3.1 Classification ability The CNN-classification with 1755395 parameters was trained on 379713 spectrograms: 126571 each for noise, tremors, and earthquakes. The model achieved 98.0% accuracy on the test data; the precision rates were 97.8%, 97.3%, and 98.8% for noise, tremor, and earthquake, respectively, and the recall rates were 96.4%, 98.4%, and 99.1%, respectively (Table S1 ). The fact that all evaluation metrics exceeded 95% demonstrated the high classification accuracy of the model. The probability distribution showed that noise, tremor, and earthquake data were positioned near the vertices of their respective categories (Fig. S5a). The result was accurate classification of 89.7%, 93.8%, and 97.6% of the noise, tremor, and earthquake, for data with prediction probabilities over 90%, respectively. The signal-to-noise ratio (SNR) is defined as \(\:\text{S}\text{N}\text{R}=20\:{\text{l}\text{o}\text{g}}_{10}\left({\text{R}\text{M}\text{S}}_{\text{t}\text{r}\text{e}\text{m}\text{o}\text{r}}/{\text{R}\text{M}\text{S}}_{\text{n}\text{o}\text{i}\text{s}\text{e}}\right)\) [dB] (Sugii et al. 2024 ), where \(\:{\text{R}\text{M}\text{S}}_{\text{t}\text{r}\text{e}\text{m}\text{o}\text{r}}\) and \(\:{\text{R}\text{M}\text{S}}_{\text{n}\text{o}\text{i}\text{s}\text{e}}\) represent the RMS amplitudes of tremors and noise over a one-minute waveform, respectively. \(\:{\text{R}\text{M}\text{S}}_{\text{n}\text{o}\text{i}\text{s}\text{e}}\) was estimated from one-minute waveform segments recorded between 01:00 and 02:00 on 23 March 2016 (JST), when no catalogued earthquakes or tremors were reported. Waveforms with a noise probability above 90% were treated as noise. For each component at each station, the lowest RMS was defined as the \(\:{\text{R}\text{M}\text{S}}_{\text{n}\text{o}\text{i}\text{s}\text{e}}\) . Tremor data were collected from 1–20 March 2016 using stations within 40 km of the epicenters in the tremor catalog. The average SNR over the components was used as the SNR for each station. When the SNR exceeded 7 dB, the indication was that the \(\:{\text{R}\text{M}\text{S}}_{\text{t}\text{r}\text{e}\text{m}\text{o}\text{r}}\) was ~ 2.2 times \(\:{\text{R}\text{M}\text{S}}_{\text{n}\text{o}\text{i}\text{s}\text{e}}\) , and the tremor detection rate remained stable, but it declined below this threshold (Fig. S5b). We also examined the relationship between epicentral distance and tremor probability using data recorded between January and September 2016 at stations within 100 km of the epicenters in the tremor catalog (Fig. S5c). The decrease of the proportion of data classified as tremors with distance to below 50% beyond 50 km was consistent with previous studies (Takahashi et al. 2021 ; Sugii et al. 2024 ). 3.2 Accuracy of tremor epicenters The CNN-regression had 1057602 trainable parameters and was trained on 76523 tremor data. A comparison of true and predicted locations for the test data showed that the CNN accurately reproduced the true locations, including the distribution of clustered tremors in eastern Shikoku (Fig. 3 a). The mean and median distances between predicted and true locations were 5.6 km and 3.2 km, respectively, and the locations of 89% of the events were within 10 km of the true locations (Fig. 3 b). This locating performance was robust for the selection of the origin of the coordinate. We evaluated the model’s ability to distinguish multiple events when the test data contained two tremors within the same window of time. When the inter-tremor distance was less than ~ 100 km, DBSCAN generally could not form two clusters (Fig. S6). The implication was that the resolution required to distinguish epicenters of multiple events was ~ 100 km, comparable to that reported by Mizuno and Ide ( 2019 ). For distances less than this threshold, the method output a single epicenter. The result was a location error of at least 20 km for the other event. In addition, the fact that many data corresponded to inter-tremor distances of less than 10 km was likely due to the close proximity of events that occurred within one minute of time (Mizuno and Ide 2019 ). For analysis of continuous data, we used the standard deviation to eliminate outliers. An analysis of the number of events with both latitude and longitude standard deviations below a given threshold, along with the mean error, revealed that lowering the threshold reduced the mean location error (Fig. 3 c). This discovery confirmed that events with greater uncertainty tended to have larger location errors. Based on this analysis, we set a standard deviation threshold of 0.065° for both latitude and longitude. The standard deviations of 95% of the test data were less than this threshold, and the mean error of those test data was 4.47 km. 3.3 Capturing spatiotemporal features of tremor activity The CNN-based approach described in section 2.2.3 was applied to the continuous data during 2012. After spatiotemporal clustering, 127524 tremor epicenters were identified, approximately 7.4 times the number identified by Mizuno and Ide ( 2019 ) during the same period (Fig. 4 a). The spatiotemporal distribution of tremors closely matched the true distribution. Even during periods of intense tremor activity, the detection performance of the proposed method was excellent (Fig. 4 b and 4 c). This result demonstrated that the method effectively overcame the tendency of low waveform similarity to lower the detection of tremors, a limitation of conventional techniques. In the Kii region, multiple, rapid tremor reversals (RTRs) were apparent during 11–16 August 2012 (Fig. 5 a and 5 b). These RTRs were consistent with the findings of Sagae et al. ( 2021 ), who were able to detect and locate tremors with high precision by using the multiple signal classification method with array data. The implication is that the resolution of the CNN-based approach proposed in this study is nearly comparable to that of Sagae et al. ( 2021 ), despite not utilizing the similarity of the waveforms of tremors and despite the much sparser distribution of stations than that of the array used by Sagae et al. ( 2021 ). 4. Conclusions The deep learning approach developed in this study, which integrates two CNNs for classification of events and estimation of the epicenters of tremors, represents a significant advancement in detecting and locating tectonic tremors in the Nankai subduction zone. The performance of the CNN-classification was excellent; it achieved over 95% accuracy of recall and precision in classifying noise, tremors, and earthquakes. The CNN-regression, with ensemble predictions from 100 models, provided precise estimates of epicenters; the median location error was 3.2 km. In addition, the proposed approach showed an enhanced ability to detect a greater number of tremors from continuous data. This improvement was especially pronounced during periods of intense tremor activity, when the proposed approach outperformed conventional techniques in detecting and locating tremors. The produced tremor catalog effectively captured the spatiotemporal features of previously documented tremors. These findings highlighted the potential of the approach based on deep learning to enhance monitoring of slow earthquakes. Abbreviations CNN convolutional neural network DBSCAN Density-Based Spatial Clustering of Applications with Noise RMS root mean square SNR signal-to-noise ratio RTR rapid tremor reversal Declarations Availability of data and materials We used catalogs reported by the Japan Meteorological Agency (https://www.data.jma.go.jp/svd/eqev/data/bulletin/hypo.html) and Mizuno and Ide (2019) (http://www-solid.eps.s.u-tokyo.ac.jp/~sloweq/). Seismograms recorded at the Hi-net (National Research Institute for Earth Science and Disaster Resilience 2019) were downloaded from the Hi-net website (https://www.hinet.bosai.go.jp). The tremor catalog for 2012 that was created in this study is available in Jinde et al. (2025). The trained CNN models are available at https://github.com/amanegeophys/TremorLocator/tree/main. Competing interests The authors declare no competing interests. Fundings AS was supported by a grant to a Research Fellowship from the Japan Society for the Promotion of Science. Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Authors’ contributions All authors contributed to the design of this study. YJ and AS conducted the analyses and all authors discussed the results. YH drafted the manuscript. All authors read and approved the final manuscript. Acknowledgements We thank the National Research Institute for Earth Science and Disaster Resilience for allowing us to use seismic waveform data. We are also grateful to Japan Meteorological Agency and Mizuno and Ide (2019) for providing catalogs. We used Python packages such as Obspy (Krischer et al. 2015) and Tensorflow (Abadi et al. 2016) in this study. The figures were produced using Generic Mapping Tools (Wessel et al. 2019) and Matplotlib (Hunter 2007). References Abadi M, Barham P, Chen J et al (2016) TensorFlow: A system for large-scale machine learning. Proc. of 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI) 16:265–283 https:// Ester M, Kriegel HP, Sander J, Xu X (1996) A density-based algorithm for discovering clusters in large spatial databases with noise. kdd 96:226–231. https://cdn.aaai.org/KDD/1996/KDD96-037.pdf Hiramatsu Y, Watanabe T, Obara K (2008) Deep low-frequency tremors as a proxy for slip monitoring at plate interface. Geophys Res Lett 35(13). https://doi.org/10.1029/2008GL034342 Hunter JD (2007) Matplotlib: a 2D graphics environment. Comput Sci Eng 9(3):90–95. https://doi.org/10.1109/MCSE.2007.55 Jinde Y, Sugii A, Hiramatsu Y (2025) Tremor epicenter catalog in 2012 in the Nankai subduction zone from CNN-based approach (from January 2012 to December 2012). [Dataset]. https://doi.org/10.5281/zenodo.15110579 . Zenodo Kato A, Obara K, Igarashi T, Tsuruoka H, Nakagawa S, Hirata N (2012) Propagation of slow slip leading up to the 2011 Mw 9.0 Tohoku-Oki earthquake. Science 335(6069):705–708. https://doi.org/10.1126/science.1215141 Kingma DP, Ba LJ (2015) Adam: a method for stochastic optimization. Paper presented at the 3rd International Conference on Learning Representations (ICLR2015), Hilton San Diego Resort-Spa, San Diego, 7–9 May 2015 Kong Q, Trugman DT, Ross ZE, Bianco MJ, Maeda B, Gerstoft P (2019) Machine learning in seismology: turning data into insights. Seismol Res Lett 90:3–14. https://doi.org/10.1785/0220180259 Kriegerowski M, Petersen GM, Vasyura-Bathke H, Ohrnberger M (2019) A deep convolutional neural network for localization of clustered earthquakes based on multistation full waveforms. Seismol Res Lett 90(2A):510–516. https://doi.org/10.1785/0220180320 Krischer L, Megies T, Barsch R, Beyreuther M, Lecocq T, Caudron C, Wassermann J (2015) ObsPy: A bridge for seismology into the scientific Python ecosystem. Comput Sci Discov 8:014003. https://doi.org/10.1088/1749-4699/8/1/014003 Kubo H, Naoi M, Kano M (2024) Recent advances in earthquake seismology using machine learning. Earth Planet Space 76:36. https://doi.org/10.1186/s40623-024-01982-0 Maeda T, Obara K (2009) Spatiotemporal distribution of seismic energy radiation from low-frequency tremor in western Shikoku, Japan. J Geophys Res Solid Earth 114:B10. https://doi.org/10.1029/2008JB006043 Matsuzawa T, Hirose H, Shibazaki B, Obara K (2010) Modeling short-and long‐term slow slip events in the seismic cycles of large subduction earthquakes. J Geophys Res Solid Earth 115:B12. https://doi.org/10.1029/2010JB007566 Mizuno N, Ide S (2019) Development of a modified envelope correlation method based on maximum-likelihood method and application to detecting and locating deep tectonic tremors in western Japan. Earth Planet Space 71:40. https://doi.org/10.1186/s40623-019-1022-x Mousavi SM, Beroza GC (2022) Deep-learning seismology. Science 377:eabm4470. https://doi.org/10.1126/science.abm4470 Nakamoto K, Hiramatsu Y, Uchide T, Imanishi K (2021) Cascading rupture of patches of high seismic energy release controls the growth process of episodic tremor and slip events. Earth Planet Space 73:59. https://doi.org/10.1186/s40623-021-01384-6 Nakano M, Sugiyama D, Hori T, Kuwatani T, Tsuboi S (2019) Discrimination of seismic signals from earthquakes and tectonic tremor by applying a convolutional neural network to running spectral images. Seismol Res Lett 90(2A):530–538. https://doi.org/10.1785/0220180279 National Research Institute for Earth Science and Disaster Resilience (2019) NIED Hi-net. National Research Institute for Earth Science and Disaster Resilience. https://doi.org/10.17598/nied.0003 Obara K (2002) Nonvolcanic deep tremor associated with subduction in southwest Japan. Science 296(5573):1679–1681. https://doi.org/10.1126/science.1070378 Obara K (2011) Characteristics and interactions between non-volcanic tremor and related slow earthquakes in the Nankai subduction zone, southwest Japan. J Geody 52(3–4):229–248. https://doi.org/10.1016/j.jog.2011.04.002 Obara K, Kato A (2016) Connecting slow earthquakes to huge earthquakes. Science 353(6296):253–257. https://doi.org/10.1126/science.aaf1512 Rogers G, Dragert H (2003) Episodic tremor and slip on the Cascadia subduction zone: The chatter of silent slip. Science 300(5627):1942–1943. https://doi.org/10.1126/science.1084783 Rouet-Leduc B, Hulbert C, McBrearty IW, Johnson PA (2020) Probing slow earthquakes with deep learning. Geophys Res Lett 47(4):e2019GL085870 https://doi.org/10.1029/2019GL085870 Ruiz S, Metois M, Fuenzalida A, Ruiz J, Leyton F, Grandin R, Vigny C, Madariaga R, Campos J (2014) Intense foreshocks and a slow slip event preceded the 2014 Iquique M W 8.1 earthquake. Science 345(6201):1165–1169. https://doi.org/10.1126/science.1256074 Sagae K, Nakahara H, Nishimura T, Imanishi K (2021) High resolution location of deep low-frequency tremors beneath the Kii Peninsula, Nankai subduction zone, Japan, using data from a dense seismic array. Geophys J Int 225(2):775–788. https://doi.org/10.1093/gji/ggab004 Shiomi K, Matsubara M, Ito Y, Obara K (2008) Simple relationship between seismic activity along Philippine Sea slab and geometry of oceanic Moho beneath southwest Japan. Geophys J Int 173:1018–1029. https://doi.org/10.1111/j.1365-246x.2008.03786.x Sugii A, Hiramatsu Y, Uchide T, Imanishi K (2024) Automated hypocenter determination of tectonic tremors in the Nankai subduction zone using convolutional neural networks combined with semblance analysis. Earth Planet Space 76:179. https://doi.org/10.1186/s40623-024-02126-0 Takahashi H, Tateiwa K, Yano K, Kano M (2021) A convolutional neural network-based classification of local earthquakes and tectonic tremors in Sanriku-oki, Japan, using S-net data. Earth Planet Space 73(186):1–10. https://doi.org/10.1186/s40623-021-01524-y Thomas AM, Inbal A, Searcy J, Shelly DR, Bürgmann R (2021) Identification of low-frequency earthquakes on the San Andreas fault with deep learning. Geophys Res Lett 48. https://doi.org/10.1029/2021GL093157 . e2021GL093157 Wessel P, Luis F, Uieda L, Scharroo R, Wobbe F, Smith WHF, Tian D (2019) The Generic Mapping Tools Version 6. G3. 20(11):5556–5564. https://doi.org/10.1029/2019GC008515 Supplementary Files graphicalabstract.png Cite Share Download PDF Status: Published Journal Publication published 22 Jul, 2025 Read the published version in Earth, Planets and Space → Version 1 posted Reviewers agreed at journal 14 May, 2025 Reviewers invited by journal 13 May, 2025 Editor assigned by journal 03 May, 2025 First submitted to journal 01 May, 2025 Editorial decision: Minor Revision 01 May, 2025 You are reading this latest preprint version 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-6526402","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":456177956,"identity":"52b8efa3-c08b-4c2c-92d8-371873c6cf67","order_by":0,"name":"Yuya Jinde","email":"","orcid":"","institution":"Kanazawa University: Kanazawa Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Yuya","middleName":"","lastName":"Jinde","suffix":""},{"id":456177957,"identity":"670e1aa4-34ea-45ce-8a46-3d1c8b29e83b","order_by":1,"name":"Amane Sugii","email":"","orcid":"","institution":"Kanazawa University: Kanazawa Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Amane","middleName":"","lastName":"Sugii","suffix":""},{"id":456177958,"identity":"1026236e-f97a-4ebc-a0b0-1b8f3e46c33c","order_by":2,"name":"Yoshihiro Hiramatsu","email":"data:image/png;base64,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","orcid":"https://orcid.org/0000-0001-9874-5059","institution":"Kanazawa University","correspondingAuthor":true,"prefix":"","firstName":"Yoshihiro","middleName":"","lastName":"Hiramatsu","suffix":""}],"badges":[],"createdAt":"2025-04-25 07:31:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6526402/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6526402/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s40623-025-02257-y","type":"published","date":"2025-07-22T15:58:17+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":82917277,"identity":"d0ee1b1b-38c2-4ffe-9415-4017b9f62cb3","added_by":"auto","created_at":"2025-05-16 16:34:09","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":718392,"visible":true,"origin":"","legend":"\u003cp\u003eDistributions of \u003cstrong\u003e(a)\u003c/strong\u003e Hi-net stations, \u003cstrong\u003e(b)\u003c/strong\u003etectonic tremors for the CNN-classification, \u003cstrong\u003e(c)\u003c/strong\u003e tectonic tremors for the CNN-regression, and \u003cstrong\u003e(d)\u003c/strong\u003e earthquakes for the CNN-classification used in this study. Dashed lines in (a)–(c) are the iso-depth contours of the subducting Philippine Sea plate (Shiomi et al. 2008).\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6526402/v1/4a5e83e0ebfb67fbc98641e7.jpeg"},{"id":82917278,"identity":"cd926235-81a9-4425-a43d-70ca467964e5","added_by":"auto","created_at":"2025-05-16 16:34:09","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":827462,"visible":true,"origin":"","legend":"\u003cp\u003eAn example of analysis of one-minute waveforms of tremor: \u003cstrong\u003e(a)\u003c/strong\u003e band-pass filtered (2–8 Hz), three-component velocity waveforms, \u003cstrong\u003e(b)\u003c/strong\u003ethe normalized spectrograms thereof, and \u003cstrong\u003e(c)\u003c/strong\u003e distributions of tremor probabilities output from the CNN-classification and RMS amplitudes, and the tremor epicenter output from (red) the CNN-regression, and (blue) that reported by Mizuno and Ide (2019).\u003c/p\u003e","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6526402/v1/e610c2d8f1d2c5f30d5cb65b.jpeg"},{"id":82918111,"identity":"73323925-3377-45cf-bb75-5dce3141e4a2","added_by":"auto","created_at":"2025-05-16 16:42:09","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":216552,"visible":true,"origin":"","legend":"\u003cp\u003eResults of the CNN-regression for the test data: \u003cstrong\u003e(a)\u003c/strong\u003espatial distributions of (red) the predicted epicenters and (blue) the true epicenters as the ground truth (Mizuno and Ide 2019), \u003cstrong\u003e(b)\u003c/strong\u003e frequency distribution of distance between predicted and true epicenters, and \u003cstrong\u003e(c)\u003c/strong\u003e(bars) ratio of predicted events below the threshold and (red circles) mean distance of errors in relation to the threshold of longitude and latitude standard deviations. The dashed line indicates a ratio of 0.95.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6526402/v1/d8bd5b4b9c3925978b3d0fdc.png"},{"id":82917280,"identity":"ba76e28c-3e68-4df9-83c7-1ed98dfb01e7","added_by":"auto","created_at":"2025-05-16 16:34:09","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":341029,"visible":true,"origin":"","legend":"\u003cp\u003eResults for continuous data during 2012: \u003cstrong\u003e(a)\u003c/strong\u003eSpatiotemporal distributions of the tremor epicenters, and the temporal variations of \u003cstrong\u003e(b)\u003c/strong\u003e cumulative numbers and \u003cstrong\u003e(c)\u003c/strong\u003e daily number of tremors. Red circles and lines are the results of this study; blue symbols the results from Mizuno and Ide (2019).\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6526402/v1/a0ccaa14b6404f3e73942e23.png"},{"id":82917284,"identity":"9c8cca88-face-4a1a-ab3c-9e47eea3d232","added_by":"auto","created_at":"2025-05-16 16:34:09","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":915436,"visible":true,"origin":"","legend":"\u003cp\u003eTremor migrations observed from \u003cstrong\u003e(a)\u003c/strong\u003e 11–16 August 2012 and \u003cstrong\u003e(b)\u003c/strong\u003e during 19:00–24:00 on 13 August 2012: (top) map view, (middle) along-strike cross-section (line A–B), and (bottom) along-dip cross-section (line C–D). Colors show the time sequence. Black lines in the cross-sections indicate the speed of migration of the RTRs.\u003c/p\u003e","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6526402/v1/6ccb547d4b9f292da80d6a75.jpeg"},{"id":87757394,"identity":"4723b544-b87b-4410-9d04-e19e11dac24b","added_by":"auto","created_at":"2025-07-28 16:10:46","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3632100,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6526402/v1/c8b7942b-158a-4cdd-9696-0eab33bddf6c.pdf"},{"id":82918113,"identity":"c80ddc0d-cea3-475c-b97a-de4fcc8c7a90","added_by":"auto","created_at":"2025-05-16 16:42:09","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":311870,"visible":true,"origin":"","legend":"","description":"","filename":"graphicalabstract.png","url":"https://assets-eu.researchsquare.com/files/rs-6526402/v1/89fe848421902b2bbd86bf72.png"}],"financialInterests":"","formattedTitle":"Enhanced deep learning approach for detecting and locating tectonic tremors in the Nankai subduction zone","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eSlow earthquakes, characterized by lower fault slips than regular earthquakes, occur mainly along plate boundaries such as the Nankai subduction zone. They are classified into geodetically detected slow-slip events (SSEs), seismically observed low-frequency earthquakes, very low-frequency earthquakes, and tectonic tremors. Short-term SSEs last weeks to months and recur every few months. Their recurrence intervals and migration patterns may change before megathrust events (e.g., Matsuzawa et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). These events may be linked to megathrust earthquakes because they occur near megathrust rupture zones (e.g., Obara and Kato \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Interactions have been suggested to occur between slow slips and megathrust earthquakes in subduction zones (Kato et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2012\u003c/span\u003e; Ruiz et al. \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2014\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTectonic tremors, weak vibrations of 1\u0026ndash;10 Hz lasting minutes to hours, often coincide with short-term SSEs to form Episodic Tremors and Slip (e.g., Rogers and Dragert \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2003\u003c/span\u003e). Because tremors serve as proxies for short-term SSEs and slip rates at plate interfaces (e.g., Hiramatsu et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2008\u003c/span\u003e; Obara \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2011\u003c/span\u003e), precise and continuous monitoring is crucial for understanding the link between slow-slip and megathrust events. However, it is challenging to detect and locate tremors because of the lack of clear P- and S-wave arrivals. The envelope correlation method (ECM) (Obara \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2002\u003c/span\u003e), which is based on waveform similarity, is commonly used for detecting and locating tremors and has been refined. Maeda and Obara (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2009\u003c/span\u003e) improved location accuracy by incorporating radiation energy, and Mizuno and Ide (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) redefined the ECM using a maximum likelihood method that enhanced the accuracy and enabled the detection of multiple sources. However, ECM-based methods often misclassify regular earthquakes and noise as tremors. High tremor activities or noise levels cause a reduction of waveform similarity that leads to missed detections.\u003c/p\u003e \u003cp\u003eAdvances in machine learning, particularly deep learning, have led to novel findings in seismology (Kong et al. \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Mousavi and Beroza \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Kubo et al. \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Convolutional neural networks (CNNs) have been widely used in earthquake analysis. Kriegerowski et al. (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) have developed a CNN-based regression model for locating hypocenters of earthquakes from seismic waveforms without phase detection. Thomas et al. (\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) have constructed a CNN for detecting low-frequency earthquakes on the Parkfield section of the San Andreas fault. CNNs have also been used to classify tremors and other events in the Nankai Trough (Nakano et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), the Japan Trench (Takahashi et al. \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), and the Cascadia subduction zone (Rouet-Leduc et al. \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). However, these CNN-based approaches face challenges of misclassifications between tremors, distant earthquakes, and noise. Recently, Sugii et al. (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) have included distant earthquakes in training data and have developed a highly accurate CNN-based approach to classify waveforms recorded by a temporal array network in the Nankai subduction zone, although their analysis has been limited to tremors around the Kii Peninsula. A new method for detecting and locating tremors using permanent stations, independent of waveform similarity and applicable across the entire Nankai subduction zone, is therefore needed.\u003c/p\u003e \u003cp\u003eIn this study, we developed a novel CNN-based approach for detecting and locating tremors in the Nankai subduction zone, consisting of a CNN for classifying events (CNN-classification) and a CNN for regression prediction of tremor locations (CNN-regression). This method outperformed conventional techniques and successfully captured previously reported spatiotemporal features of tremors.\u003c/p\u003e"},{"header":"2. Data and Methods","content":"\u003cp\u003eWe used three-component velocity waveforms recorded at 129 Hi-net stations operated by the National Research Institute for Earth Science and Disaster Resilience around the Nankai subduction zone from January 2008 to September 2016 (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Classification of events\u003c/h2\u003e \u003cdiv id=\"Sec4\" class=\"Section3\"\u003e \u003ch2\u003e2.1.1 Datasets for the CNN-classification\u003c/h2\u003e \u003cp\u003eWe constructed a dataset with three categories: noise, tremors, and earthquakes. Tremor data were identified by reference to the origin times and epicenters in the tremor catalog of Mizuno and Ide (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Periods with at least five tremors per hour were identified, and waveforms were visually inspected. Only confirmed tremor waveforms were included. Waveforms from stations over 40 km from the epicenter were excluded to maintain classification accuracy (Sugii et al. \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). For unlisted times, epicentral distances were estimated based on the average locations of tremors that occurred within one hour.\u003c/p\u003e \u003cp\u003eEarthquake data, such as hypocenters and arrival times, were obtained from catalogs reported by the Japan Meteorological Agency. To preserve classification accuracy, events recorded at any of the 129 stations were included, except for those shallower than 50 km with magnitudes less than 1 or deeper than 50 km with magnitudes less than 3. The start times of one-minute windows were randomly shifted by up to 45 seconds preceding the arrival times of the P-wave to enhance model robustness, and one-minute waveforms were extracted.\u003c/p\u003e \u003cp\u003eWaveforms that were not classified as tremors or earthquakes were treated to be noise. For noise data selection, visual inspections at multiple stations ensured the exclusion of waveforms with abnormal signals or features indicating tremors or earthquakes.\u003c/p\u003e \u003cp\u003eTo balance the data, we applied random sampling and excluded waveforms with missing components. Each category (noise, tremors, and earthquakes) consequently obtained 126571 spectrograms. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb and \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ed show the distributions of the epicenters of tremors and earthquakes used for the CNN-classification. Training and validation data spanned the period from January 2008 to December 2015 and consisted of 104091 spectrograms per category. Ninety percent of these spectrograms were randomly chosen for training, and the remainder were used for validation. The test data consisted of 22480 spectrograms per category and covered the period from January to September 2016 (Figs. S1 and S2).\u003c/p\u003e \u003cp\u003eWe prepared input data for the CNN-classification following Sugii et al. (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Spectrograms were generated from 2\u0026ndash;8 Hz bandpass-filtered, three-component waveforms of one-minute duration, sampled at 5-second intervals with a lag of 1.19 seconds (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea and \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb). Each spectrogram consisted of 52 \u0026times; 52 pixels and covered 0\u0026ndash;10 Hz on a linear scale. Amplitude spectra were normalized by subtracting the mean and dividing by the standard deviation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e2.1.2 Architecture and training of CNN-classification\u003c/h2\u003e \u003cp\u003eThe CNN-classification followed the architecture of Sugii et al. (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), which includes three convolutional layers with 3 \u0026times; 3 filters, two MaxPooling layers, two dropout layers, and one softmax layer, along with fully connected layers of 32 and 3 nodes (Fig. S3a). Rectified Linear Units were used as the activation function. The batch size and learning rate were set to 64 and 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e, respectively, based on the most accurate model of Sugii et al. (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). We used the Adam algorithm (Kingma and Ba \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) to apply the categorical cross-entropy loss function.\u003c/p\u003e \u003cp\u003eTraining stopped if the validation loss did not improve during four consecutive epochs or after 500 epochs. The model with the lowest validation loss was used for testing. For each inputted one-minute spectrogram, the CNN output the probabilities of classifying the waveform as noise, tremor, or earthquake.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Locating tremor epicenters\u003c/h2\u003e \u003cp\u003eThe spatial distribution of amplitudes facilitated locating tremors (e.g., Maeda and Obara \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). Furthermore, the tremor probabilities output by the CNN-classification provided information that facilitated estimating tremor epicenters because they tended to be high and form clusters near the epicenter (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec). In this study, we used tremor probabilities for waveform preprocessing and developed the CNN-regression that determined tremor epicenters based on amplitude data.\u003c/p\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.2.1 Dataset for CNN-regression\u003c/h2\u003e \u003cp\u003eWe applied the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm (Ester et al. \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e1996\u003c/span\u003e) to detect clusters of stations with high tremor probabilities based on the CNN-classification. For cataloged times, clustering using DBSCAN was performed on stations with tremor probabilities of 0.9 or higher, using ε\u0026thinsp;=\u0026thinsp;0.5\u0026deg; and MinPts\u0026thinsp;=\u0026thinsp;3, where ε is the radius of the neighborhood, and MinPts is the minimum number of stations. These parameters were determined through trial and error. We selected events with clusters for training, validation, and testing. If multiple clusters were formed simultaneously, separate data were created for each cluster. We excluded events with fewer than three stations, at which tremor probabilities were 0.9 or greater, within 0.5\u0026deg; of the epicenters. In addition, we excluded data from 2012 to conduct the continuous data analysis. The final dataset consisted of 116136 events (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ec) divided into training (76523: January 2008 to December 2013 except for 2012), validation (15573: January 2014 to December 2014), and test (24040: January 2015 to September 2016) data (Fig. S4).\u003c/p\u003e \u003cp\u003eThe CNN-regression used RMS amplitudes from bandpass-filtered (2\u0026ndash;8 Hz), one-minute waveforms. Epicenters were taken from the tremor catalog of Mizuno and Ide (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) and served as ground truth. The input data consisted of a 3 \u0026times; 129 matrix that represented the root mean square (RMS) amplitude of the north\u0026ndash;south, east\u0026ndash;west, and vertical components of the waveforms at 129 stations. To reduce the impact of abnormal amplitudes, values exceeding three standard deviations above the mean RMS amplitude were corrected to 10\u003csup\u003e\u0026minus;\u0026thinsp;9\u003c/sup\u003e m/s, comparable to the average noise level. Similarly, amplitude corrections were applied to stations with tremor probabilities less than 0.9, to those classified as non-clusters by DBSCAN, and to those assigned to non-target clusters. Finally, the amplitudes were standardized for each component by subtracting the mean and dividing by the standard deviation to ensure uniform scaling.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.2.2 Architecture and training of CNN-regression\u003c/h2\u003e \u003cp\u003eThe CNN-regression consisted of two convolutional layers with 3 \u0026times; 2 filters, followed by fully connected layers with 128, 64, and 2 nodes (Fig. S3b). The omission of pooling layers to preserve spatial information enabled the model to implicitly learn the locations of stations. Epicenters from the catalog were converted into latitude and longitude differences relative to an origin at 34.1\u0026deg;N, 135\u0026deg;E for stable and efficient training. The output layer predicted the latitude and longitude relative to this origin.\u003c/p\u003e \u003cp\u003eThe CNN-regression was trained using the mean absolute error as the loss function and the Adam algorithm with a batch size of 32 and a learning rate of 5 \u0026times; 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e. Training stopped if the loss did not improve for 30 consecutive epochs or after 400 epochs.\u003c/p\u003e \u003cp\u003eWe used ensemble learning with 100 models to assess prediction uncertainty. Each model with randomly initialized parameters was trained to minimize validation losses. The average and standard deviation of those predictions were treated as the final model prediction and uncertainty, respectively.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e2.2.3 Application to continuous data\u003c/h2\u003e \u003cp\u003eFor continuous data, spectrograms at one-minute intervals for 129 stations were input into the CNN-classification. If DBSCAN formed clusters of stations with tremor probabilities of 0.9 or greater, the data with their RMS amplitude corrections were input into the CNN-regression to locate epicenters. When multiple clusters were detected, the epicenter was estimated for each cluster.\u003c/p\u003e \u003cp\u003eEvents exceeding the uncertainty threshold determined from the test data (see section \u003cspan refid=\"Sec12\" class=\"InternalRef\"\u003e3.2\u003c/span\u003e for details) were excluded as outliers. In addition, events were excluded if they were detected at fewer than three stations, at which tremor probabilities were 0.9 or greater, located within a 0.5\u0026deg; square of the estimated epicenter. Events were also excluded as isolated if no adjacent events were detected within 5 km during a 12-hour window before or after the event (Nakamoto et al. \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"3. Results and Discussion","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Classification ability\u003c/h2\u003e \u003cp\u003eThe CNN-classification with 1755395 parameters was trained on 379713 spectrograms: 126571 each for noise, tremors, and earthquakes. The model achieved 98.0% accuracy on the test data; the precision rates were 97.8%, 97.3%, and 98.8% for noise, tremor, and earthquake, respectively, and the recall rates were 96.4%, 98.4%, and 99.1%, respectively (Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). The fact that all evaluation metrics exceeded 95% demonstrated the high classification accuracy of the model. The probability distribution showed that noise, tremor, and earthquake data were positioned near the vertices of their respective categories (Fig. S5a). The result was accurate classification of 89.7%, 93.8%, and 97.6% of the noise, tremor, and earthquake, for data with prediction probabilities over 90%, respectively.\u003c/p\u003e \u003cp\u003eThe signal-to-noise ratio (SNR) is defined as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{S}\\text{N}\\text{R}=20\\:{\\text{l}\\text{o}\\text{g}}_{10}\\left({\\text{R}\\text{M}\\text{S}}_{\\text{t}\\text{r}\\text{e}\\text{m}\\text{o}\\text{r}}/{\\text{R}\\text{M}\\text{S}}_{\\text{n}\\text{o}\\text{i}\\text{s}\\text{e}}\\right)\\)\u003c/span\u003e\u003c/span\u003e [dB] (Sugii et al. \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{R}\\text{M}\\text{S}}_{\\text{t}\\text{r}\\text{e}\\text{m}\\text{o}\\text{r}}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{R}\\text{M}\\text{S}}_{\\text{n}\\text{o}\\text{i}\\text{s}\\text{e}}\\)\u003c/span\u003e\u003c/span\u003e represent the RMS amplitudes of tremors and noise over a one-minute waveform, respectively. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{R}\\text{M}\\text{S}}_{\\text{n}\\text{o}\\text{i}\\text{s}\\text{e}}\\)\u003c/span\u003e\u003c/span\u003e was estimated from one-minute waveform segments recorded between 01:00 and 02:00 on 23 March 2016 (JST), when no catalogued earthquakes or tremors were reported. Waveforms with a noise probability above 90% were treated as noise. For each component at each station, the lowest RMS was defined as the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{R}\\text{M}\\text{S}}_{\\text{n}\\text{o}\\text{i}\\text{s}\\text{e}}\\)\u003c/span\u003e\u003c/span\u003e. Tremor data were collected from 1\u0026ndash;20 March 2016 using stations within 40 km of the epicenters in the tremor catalog. The average SNR over the components was used as the SNR for each station. When the SNR exceeded 7 dB, the indication was that the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{R}\\text{M}\\text{S}}_{\\text{t}\\text{r}\\text{e}\\text{m}\\text{o}\\text{r}}\\)\u003c/span\u003e\u003c/span\u003e was ~\u0026thinsp;2.2 times \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{R}\\text{M}\\text{S}}_{\\text{n}\\text{o}\\text{i}\\text{s}\\text{e}}\\)\u003c/span\u003e\u003c/span\u003e, and the tremor detection rate remained stable, but it declined below this threshold (Fig. S5b).\u003c/p\u003e \u003cp\u003eWe also examined the relationship between epicentral distance and tremor probability using data recorded between January and September 2016 at stations within 100 km of the epicenters in the tremor catalog (Fig. S5c). The decrease of the proportion of data classified as tremors with distance to below 50% beyond 50 km was consistent with previous studies (Takahashi et al. \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Sugii et al. \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Accuracy of tremor epicenters\u003c/h2\u003e \u003cp\u003eThe CNN-regression had 1057602 trainable parameters and was trained on 76523 tremor data. A comparison of true and predicted locations for the test data showed that the CNN accurately reproduced the true locations, including the distribution of clustered tremors in eastern Shikoku (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea). The mean and median distances between predicted and true locations were 5.6 km and 3.2 km, respectively, and the locations of 89% of the events were within 10 km of the true locations (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb). This locating performance was robust for the selection of the origin of the coordinate.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eWe evaluated the model\u0026rsquo;s ability to distinguish multiple events when the test data contained two tremors within the same window of time. When the inter-tremor distance was less than ~\u0026thinsp;100 km, DBSCAN generally could not form two clusters (Fig. S6). The implication was that the resolution required to distinguish epicenters of multiple events was ~\u0026thinsp;100 km, comparable to that reported by Mizuno and Ide (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). For distances less than this threshold, the method output a single epicenter. The result was a location error of at least 20 km for the other event. In addition, the fact that many data corresponded to inter-tremor distances of less than 10 km was likely due to the close proximity of events that occurred within one minute of time (Mizuno and Ide \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFor analysis of continuous data, we used the standard deviation to eliminate outliers. An analysis of the number of events with both latitude and longitude standard deviations below a given threshold, along with the mean error, revealed that lowering the threshold reduced the mean location error (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ec). This discovery confirmed that events with greater uncertainty tended to have larger location errors. Based on this analysis, we set a standard deviation threshold of 0.065\u0026deg; for both latitude and longitude. The standard deviations of 95% of the test data were less than this threshold, and the mean error of those test data was 4.47 km.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Capturing spatiotemporal features of tremor activity\u003c/h2\u003e \u003cp\u003eThe CNN-based approach described in section \u003cspan refid=\"Sec9\" class=\"InternalRef\"\u003e2.2.3\u003c/span\u003e was applied to the continuous data during 2012. After spatiotemporal clustering, 127524 tremor epicenters were identified, approximately 7.4 times the number identified by Mizuno and Ide (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) during the same period (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea). The spatiotemporal distribution of tremors closely matched the true distribution. Even during periods of intense tremor activity, the detection performance of the proposed method was excellent (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb and \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ec). This result demonstrated that the method effectively overcame the tendency of low waveform similarity to lower the detection of tremors, a limitation of conventional techniques.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn the Kii region, multiple, rapid tremor reversals (RTRs) were apparent during 11\u0026ndash;16 August 2012 (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea and \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb). These RTRs were consistent with the findings of Sagae et al. (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), who were able to detect and locate tremors with high precision by using the multiple signal classification method with array data. The implication is that the resolution of the CNN-based approach proposed in this study is nearly comparable to that of Sagae et al. (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), despite not utilizing the similarity of the waveforms of tremors and despite the much sparser distribution of stations than that of the array used by Sagae et al. (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Conclusions","content":"\u003cp\u003eThe deep learning approach developed in this study, which integrates two CNNs for classification of events and estimation of the epicenters of tremors, represents a significant advancement in detecting and locating tectonic tremors in the Nankai subduction zone. The performance of the CNN-classification was excellent; it achieved over 95% accuracy of recall and precision in classifying noise, tremors, and earthquakes. The CNN-regression, with ensemble predictions from 100 models, provided precise estimates of epicenters; the median location error was 3.2 km. In addition, the proposed approach showed an enhanced ability to detect a greater number of tremors from continuous data. This improvement was especially pronounced during periods of intense tremor activity, when the proposed approach outperformed conventional techniques in detecting and locating tremors. The produced tremor catalog effectively captured the spatiotemporal features of previously documented tremors. These findings highlighted the potential of the approach based on deep learning to enhance monitoring of slow earthquakes.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCNN\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003econvolutional neural network\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDBSCAN\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDensity-Based Spatial Clustering of Applications with Noise\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eRMS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eroot mean square\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSNR\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003esignal-to-noise ratio\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eRTR\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003erapid tremor reversal\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe used catalogs reported by the Japan Meteorological Agency (https://www.data.jma.go.jp/svd/eqev/data/bulletin/hypo.html) and Mizuno and Ide (2019) (http://www-solid.eps.s.u-tokyo.ac.jp/~sloweq/). Seismograms recorded at the Hi-net (National Research Institute for Earth Science and Disaster Resilience 2019) were downloaded from the Hi-net website (https://www.hinet.bosai.go.jp). The tremor catalog for 2012 that was created in this study is available in Jinde et al. (2025). The trained CNN models are available at https://github.com/amanegeophys/TremorLocator/tree/main.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFundings\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAS was\u0026nbsp;supported by a grant to a Research Fellowship from the Japan Society for the Promotion of Science.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors’ contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors contributed to the design of this study. YJ and AS conducted the analyses and all authors discussed the results. YH drafted the manuscript. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank the National Research Institute for Earth Science and Disaster Resilience for allowing us to use seismic waveform data. We are also grateful to Japan Meteorological Agency and Mizuno and Ide (2019) for providing catalogs. We used Python packages such as Obspy (Krischer et al. 2015) and Tensorflow (Abadi et al. 2016) in this study. The figures were produced using Generic Mapping Tools (Wessel et al. 2019) and Matplotlib (Hunter 2007).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAbadi M, Barham P, Chen J et al (2016) TensorFlow: A system for large-scale machine learning. Proc. of 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI) 16:265\u0026ndash;283\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ehttps://\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e\u003c/span\u003e\u003cspan address=\"http://www.usenix.org/conference/osdi16/technical-sessions/presentation/abadi\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEster M, Kriegel HP, Sander J, Xu X (1996) A density-based algorithm for discovering clusters in large spatial databases with noise. kdd 96:226\u0026ndash;231. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://cdn.aaai.org/KDD/1996/KDD96-037.pdf\u003c/span\u003e\u003cspan address=\"https://cdn.aaai.org/KDD/1996/KDD96-037.pdf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHiramatsu Y, Watanabe T, Obara K (2008) Deep low-frequency tremors as a proxy for slip monitoring at plate interface. Geophys Res Lett 35(13). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1029/2008GL034342\u003c/span\u003e\u003cspan address=\"10.1029/2008GL034342\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHunter JD (2007) Matplotlib: a 2D graphics environment. Comput Sci Eng 9(3):90\u0026ndash;95. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/MCSE.2007.55\u003c/span\u003e\u003cspan address=\"10.1109/MCSE.2007.55\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJinde Y, Sugii A, Hiramatsu Y (2025) Tremor epicenter catalog in 2012 in the Nankai subduction zone from CNN-based approach (from January 2012 to December 2012). [Dataset]. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5281/zenodo.15110579\u003c/span\u003e\u003cspan address=\"10.5281/zenodo.15110579\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Zenodo\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKato A, Obara K, Igarashi T, Tsuruoka H, Nakagawa S, Hirata N (2012) Propagation of slow slip leading up to the 2011 Mw 9.0 Tohoku-Oki earthquake. Science 335(6069):705\u0026ndash;708. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1126/science.1215141\u003c/span\u003e\u003cspan address=\"10.1126/science.1215141\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKingma DP, Ba LJ (2015) Adam: a method for stochastic optimization. Paper presented at the 3rd International Conference on Learning Representations (ICLR2015), Hilton San Diego Resort-Spa, San Diego, 7\u0026ndash;9 May 2015\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKong Q, Trugman DT, Ross ZE, Bianco MJ, Maeda B, Gerstoft P (2019) Machine learning in seismology: turning data into insights. Seismol Res Lett 90:3\u0026ndash;14. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1785/0220180259\u003c/span\u003e\u003cspan address=\"10.1785/0220180259\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKriegerowski M, Petersen GM, Vasyura-Bathke H, Ohrnberger M (2019) A deep convolutional neural network for localization of clustered earthquakes based on multistation full waveforms. Seismol Res Lett 90(2A):510\u0026ndash;516. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1785/0220180320\u003c/span\u003e\u003cspan address=\"10.1785/0220180320\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKrischer L, Megies T, Barsch R, Beyreuther M, Lecocq T, Caudron C, Wassermann J (2015) ObsPy: A bridge for seismology into the scientific Python ecosystem. Comput Sci Discov 8:014003. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1088/1749-4699/8/1/014003\u003c/span\u003e\u003cspan address=\"10.1088/1749-4699/8/1/014003\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKubo H, Naoi M, Kano M (2024) Recent advances in earthquake seismology using machine learning. Earth Planet Space 76:36. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s40623-024-01982-0\u003c/span\u003e\u003cspan address=\"10.1186/s40623-024-01982-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMaeda T, Obara K (2009) Spatiotemporal distribution of seismic energy radiation from low-frequency tremor in western Shikoku, Japan. J Geophys Res Solid Earth 114:B10. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1029/2008JB006043\u003c/span\u003e\u003cspan address=\"10.1029/2008JB006043\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMatsuzawa T, Hirose H, Shibazaki B, Obara K (2010) Modeling short-and long‐term slow slip events in the seismic cycles of large subduction earthquakes. J Geophys Res Solid Earth 115:B12. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1029/2010JB007566\u003c/span\u003e\u003cspan address=\"10.1029/2010JB007566\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMizuno N, Ide S (2019) Development of a modified envelope correlation method based on maximum-likelihood method and application to detecting and locating deep tectonic tremors in western Japan. Earth Planet Space 71:40. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s40623-019-1022-x\u003c/span\u003e\u003cspan address=\"10.1186/s40623-019-1022-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMousavi SM, Beroza GC (2022) Deep-learning seismology. Science 377:eabm4470. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1126/science.abm4470\u003c/span\u003e\u003cspan address=\"10.1126/science.abm4470\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNakamoto K, Hiramatsu Y, Uchide T, Imanishi K (2021) Cascading rupture of patches of high seismic energy release controls the growth process of episodic tremor and slip events. Earth Planet Space 73:59. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s40623-021-01384-6\u003c/span\u003e\u003cspan address=\"10.1186/s40623-021-01384-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNakano M, Sugiyama D, Hori T, Kuwatani T, Tsuboi S (2019) Discrimination of seismic signals from earthquakes and tectonic tremor by applying a convolutional neural network to running spectral images. Seismol Res Lett 90(2A):530\u0026ndash;538. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1785/0220180279\u003c/span\u003e\u003cspan address=\"10.1785/0220180279\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNational Research Institute for Earth Science and Disaster Resilience (2019) NIED Hi-net. National Research Institute for Earth Science and Disaster Resilience. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.17598/nied.0003\u003c/span\u003e\u003cspan address=\"10.17598/nied.0003\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eObara K (2002) Nonvolcanic deep tremor associated with subduction in southwest Japan. Science 296(5573):1679\u0026ndash;1681. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1126/science.1070378\u003c/span\u003e\u003cspan address=\"10.1126/science.1070378\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eObara K (2011) Characteristics and interactions between non-volcanic tremor and related slow earthquakes in the Nankai subduction zone, southwest Japan. J Geody 52(3\u0026ndash;4):229\u0026ndash;248. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jog.2011.04.002\u003c/span\u003e\u003cspan address=\"10.1016/j.jog.2011.04.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eObara K, Kato A (2016) Connecting slow earthquakes to huge earthquakes. Science 353(6296):253\u0026ndash;257. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1126/science.aaf1512\u003c/span\u003e\u003cspan address=\"10.1126/science.aaf1512\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRogers G, Dragert H (2003) Episodic tremor and slip on the Cascadia subduction zone: The chatter of silent slip. Science 300(5627):1942\u0026ndash;1943. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1126/science.1084783\u003c/span\u003e\u003cspan address=\"10.1126/science.1084783\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRouet-Leduc B, Hulbert C, McBrearty IW, Johnson PA (2020) Probing slow earthquakes with deep learning. Geophys Res Lett 47(4):e2019GL085870\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1029/2019GL085870\u003c/span\u003e\u003cspan address=\"10.1029/2019GL085870\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRuiz S, Metois M, Fuenzalida A, Ruiz J, Leyton F, Grandin R, Vigny C, Madariaga R, Campos J (2014) Intense foreshocks and a slow slip event preceded the 2014 Iquique M\u003csub\u003eW\u003c/sub\u003e 8.1 earthquake. Science 345(6201):1165\u0026ndash;1169. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1126/science.1256074\u003c/span\u003e\u003cspan address=\"10.1126/science.1256074\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSagae K, Nakahara H, Nishimura T, Imanishi K (2021) High resolution location of deep low-frequency tremors beneath the Kii Peninsula, Nankai subduction zone, Japan, using data from a dense seismic array. Geophys J Int 225(2):775\u0026ndash;788. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/gji/ggab004\u003c/span\u003e\u003cspan address=\"10.1093/gji/ggab004\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShiomi K, Matsubara M, Ito Y, Obara K (2008) Simple relationship between seismic activity along Philippine Sea slab and geometry of oceanic Moho beneath southwest Japan. Geophys J Int 173:1018\u0026ndash;1029. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1111/j.1365-246x.2008.03786.x\u003c/span\u003e\u003cspan address=\"10.1111/j.1365-246x.2008.03786.x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSugii A, Hiramatsu Y, Uchide T, Imanishi K (2024) Automated hypocenter determination of tectonic tremors in the Nankai subduction zone using convolutional neural networks combined with semblance analysis. Earth Planet Space 76:179. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s40623-024-02126-0\u003c/span\u003e\u003cspan address=\"10.1186/s40623-024-02126-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTakahashi H, Tateiwa K, Yano K, Kano M (2021) A convolutional neural network-based classification of local earthquakes and tectonic tremors in Sanriku-oki, Japan, using S-net data. Earth Planet Space 73(186):1\u0026ndash;10. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s40623-021-01524-y\u003c/span\u003e\u003cspan address=\"10.1186/s40623-021-01524-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThomas AM, Inbal A, Searcy J, Shelly DR, B\u0026uuml;rgmann R (2021) Identification of low-frequency earthquakes on the San Andreas fault with deep learning. Geophys Res Lett 48. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1029/2021GL093157\u003c/span\u003e\u003cspan address=\"10.1029/2021GL093157\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. e2021GL093157\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWessel P, Luis F, Uieda L, Scharroo R, Wobbe F, Smith WHF, Tian D (2019) The Generic Mapping Tools Version 6. G3. 20(11):5556\u0026ndash;5564. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1029/2019GC008515\u003c/span\u003e\u003cspan address=\"10.1029/2019GC008515\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"earth-planets-and-space","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"epsp","sideBox":"Learn more about [Earth, Planets and Space](http://earth-planets-space.springeropen.com)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/epsp/default.aspx","title":"Earth, Planets and Space","twitterHandle":"@SpringerOpen","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"convolutional neural network, slow slip, tremor migration, rapid tremor reversal","lastPublishedDoi":"10.21203/rs.3.rs-6526402/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6526402/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eTectonic tremors are key indicators of slow-slip phenomena, and detecting them accurately is a challenging task. Conventional techniques often fail to detect tremors during periods of intense tremor activity. We present here a deep learning approach for detecting and locating tremors in the Nankai subduction zone that is more effective than conventional techniques. We utilized two convolutional neural networks (CNNs): a CNN for classification of seismic waveforms into noise, tremors, or earthquakes, and a CNN for regression prediction of tremor epicenters from amplitude data. The accuracy, recall, and precision of the CNN for classification all exceeded 95%. The CNN for regression used ensemble predictions to produce estimates of tremor locations with a median error of 3.2 km. When this approach was applied to continuous data, it successfully mapped key features of tremor activity and improved the detection and location of tremors, especially during high-activity periods.\u003c/p\u003e","manuscriptTitle":"Enhanced deep learning approach for detecting and locating tectonic tremors in the Nankai subduction zone","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-16 16:34:04","doi":"10.21203/rs.3.rs-6526402/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"","date":"2025-05-14T08:25:45+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-13T21:47:10+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-05-03T21:24:53+00:00","index":"","fulltext":""},{"type":"submitted","content":"Earth, Planets and Space","date":"2025-05-01T22:09:10+00:00","index":"","fulltext":""},{"type":"decision","content":"Minor Revision","date":"2025-05-01T20:59:22+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"earth-planets-and-space","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"epsp","sideBox":"Learn more about [Earth, Planets and Space](http://earth-planets-space.springeropen.com)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/epsp/default.aspx","title":"Earth, Planets and Space","twitterHandle":"@SpringerOpen","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"831caccb-4a8f-4d58-9178-7c2322782844","owner":[],"postedDate":"May 16th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-07-28T16:09:11+00:00","versionOfRecord":{"articleIdentity":"rs-6526402","link":"https://doi.org/10.1186/s40623-025-02257-y","journal":{"identity":"earth-planets-and-space","isVorOnly":false,"title":"Earth, Planets and Space"},"publishedOn":"2025-07-22 15:58:17","publishedOnDateReadable":"July 22nd, 2025"},"versionCreatedAt":"2025-05-16 16:34:04","video":"","vorDoi":"10.1186/s40623-025-02257-y","vorDoiUrl":"https://doi.org/10.1186/s40623-025-02257-y","workflowStages":[]},"version":"v1","identity":"rs-6526402","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6526402","identity":"rs-6526402","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}