PhaseNeXt: Neural phase picker trained on 20-year records to process the JMA-unified data set

PhaseNeXt: Neural phase picker trained on 20-year records to process the JMA-unified data set | 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 PhaseNeXt: Neural phase picker trained on 20-year records to process the JMA-unified data set Makoto Naoi, Kengo Shimojo, Koji Tamaribuchi This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8174647/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract To develop a higher-quality seismic event catalog from Japan’s routine seismic observations, which have been continuously recorded at approximately 2000 stations, we trained deep-learning-based phase pickers using 20 years of arrival-time data read by the Japan Meteorological Agency (JMA). To enhance performance, we developed a new model, PhaseNeXt, by incorporating techniques proven effective in the field of computer vision—particularly in semantic segmentation—into PhaseNet, one of the most widely used neural phase pickers. The resulting model adopts an architecture inspired by DeepLab v3+, connecting parameter-efficient ConvNeXt blocks through residual connections, which mitigate vanishing gradients and allow scalable adaptation to larger training data sets. Furthermore, by including automatically read or briefly reviewed arrival-time readings that had not undergone detailed manual inspection in the training process, we demonstrated improved performance for small earthquakes. Using these insights, we trained three deep neural network models with different parameter sizes on 25 million waveforms associated with events listed in the 2002–2023 JMA unified catalog. When integrated into the current JMA workflow, the best-performing model detected approximately 3.5 times more events than those listed in the JMA catalog while using nearly the same number of arrival-time readings. Neural phase picker Deep learning Hypocenter location Earthquake catalog development 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 Figure 13 1 Introduction In developing an earthquake catalog from seismic waveform records, the performance of phase picking—that is, reading arrival times of body waves—is a crucial factor. In recent years, studies on deep-learning-based phase picking (neural phase picking) have become increasingly active, demonstrating performance superior to that of conventional automatic processing methods (Kubo et al. 2024 ). Models trained on large data sets have been released and shared publicly (e.g., Naoi et al. 2024 ; Suzuki et al. 2025 ) and are now being employed in seismic activity analyses (Kato 2024 ). Although these models exhibit a certain level of generalization and perform reasonably well even on data from regions different from those of the training data set (Hara et al. 2019 ), their performance can be further improved when trained with data obtained from the target region and its specific observation stations. Therefore, for data sets that have a large number of users, it is desirable to develop and share models trained directly on those data (Naoi et al. 2024 ). In Japan, following the 1995 Kobe earthquake (Mw 6.9), seismic data that had previously been recorded independently by many institutions were integrated and shared. Since then, the Japan Meteorological Agency (JMA), in collaboration with related organizations, has been responsible for reading arrival times and compiling a unified earthquake catalog (hereinafter referred to as the JMA unified catalog). With the introduction of the Hi-net network (High Sensitivity Seismograph Network in Japan; NIED 2019a) in 2000, earthquake detection capability in Japan has greatly improved, and the seismic network has continued to expand coverage to offshore areas (Aoi et al. 2020 ). At present, continuous waveform data from approximately 2000 stations are routinely recorded, and both the catalog and associated arrival-time data are publicly available on the JMA website (JMA 2025). Naoi et al. ( 2024 ) retrained PhaseNet (Zhu and Beroza 2019 )—one of the most widely used neural phase pickers—using 6.1 million arrival-time readings that had been manually reviewed in detail from the JMA unified catalog covering 2014–2021. The number of training waveforms was roughly ten times greater than that used in the original PhaseNet, which had been trained on California data (Zhu and Beroza 2019 ). The model developed by Naoi et al. ( 2024 ) exhibited significantly better performance for Japan’s routine data than the original PhaseNet. They also developed an enlarged model, PhaseNetWC, in which the number of convolutional-layer channels was doubled, resulting in higher performance. The learning curves from these training processes showed no sign of overfitting, suggesting that models with a larger number of parameters and higher expressiveness—that is, a greater ability to approximate complex functions—could potentially achieve better performance. In this study, we developed a neural phase picker that achieves higher performance for Japan’s routine seismic observation data (hereafter referred to as the JMA unified data set). This study addressed the following four tasks: (1) modifying the model architecture to achieve a more parameter-efficient and scalable design; (2) examining preprocessing (preconditioning) methods for input waveforms; (3) demonstrating that incorporating arrival-time readings that had not undergone detailed manual review can improve performance for small-magnitude events; and (4) training three models with different parameter sizes using 20 years of data (2002–2012 and 2014–2022), based on the insights from (1) and (3), and evaluating their performance relative to the models developed by Naoi et al. ( 2024 ). The trained models are publicly available in the authors’ GitHub repository. 2 Data set preparation 2.1 Data extraction from the JMA unified catalog In constructing the JMA unified catalog, all arrival times used for hypocenter determination were processed manually until March 2016. In April 2016, the JMA introduced an automatic processing system based on the Phase combination Forward search (PF) method (Tamaribuchi et al. 2016 ; Tamaribuchi 2018 ). Since then, only events exceeding regionally defined magnitude thresholds have undergone detailed manual inspection (hereinafter referred to as a full manual review) and have been manually repicked when necessary. The threshold is set to M = 1.7 for shallow inland earthquakes and increases with distance from land, reaching up to M = 3.5 for oceanic earthquakes. For events with magnitudes below these thresholds, automatically processed or briefly reviewed results are published. Until March 21, 2018, full reviews for subthreshold events continued only in the Tokai region, but after that, this policy was applied nationwide. Detailed explanations of these procedures are provided in Naoi et al. ( 2024 ) and in the JMA (2025) User’s Guide. In the JMA unified catalog, various labels are assigned to each event and to each arrival-time record (JMA 2025), allowing data to be selectively extracted. Naoi et al. ( 2024 ) selected training data based on the data information flags, which are assigned to each arrival-time reading at individual stations. Each flag consists of a single alphabetic character, with 48 possible patterns in total. Of these, the 24 capital letters denote high-confidence readings that were fully reviewed. Before the introduction of the PF method, all readings were labeled with capital letters, and lowercase flags—representing lower-confidence readings—were introduced thereafter. Naoi et al. ( 2024 ) trained PhaseNet and PhaseNetWC using only data associated with capital-letter flags. However, for networks newly deployed after the introduction of the PF method—such as S-net (the Seafloor Observation Network for Earthquakes and Tsunamis along the Japan Trench; NIED 2019b)—high-confidence picks were unavailable for small events below the regional magnitude thresholds, possibly resulting in degraded model performance for such events. This issue and our approach to address it are discussed in Sections 2.3 and 4.2 . Data selection based on the data information flags can alternatively be performed more simply using the hypocenter determination flag assigned to each event. This label consists of eight alphabetic characters (Table 1 ). In this study, we analyzed only events labeled K, which represent the highest quality, and those labeled A or k, which indicate the next-highest quality (JMA 2025). The frequency-magnitude distribution (FMD) of events with these labels is shown in Fig. 1 . Before April 2016, no events with k or A labels existed. After the introduction of the PF method in April 2016, the number of small events (M < 1.7) labeled K decreased markedly and became nearly absent after March 22, 2018. Events labeled k or A are mostly limited to small magnitudes. Table 1 Hypocenter determination flags (adapted from JMA 2025). Flag description K High-precision hypocenters (Manual, closely examined) S Low-precision hypocenters (Manual, closely examined) k Middle-precision hypocenters (Manual) s Low-precision hypocenters (Manual) A Middle-precision hypocenters (Auto) a Low-precision hypocenters (Auto) N Undetermined or not accepted or fixed hypocenters F Far field In principle, arrival-time readings of events labeled K are accompanied by capital-letter data information flags. However, during a short exceptional period in late November 2016, some K-labeled events were associated with lowercase flags. This occurred because aftershocks following the Mw 7.4 Fukushima-oki earthquake on November 22, 2016, were processed using a procedure different from the routine one. These exceptions account for only 0.04% of all K-labeled data. The k and A labels were introduced after the implementation of the PF method. For events below the threshold magnitude for a full manual review, k indicates that a simplified, brief review was conducted, whereas A indicates that the automatic result derived by the PF method was retained without manual inspection. The arrival-time readings of events with k or A labels are generally accompanied by lowercase data information flags. However, a small number of exceptions exist: some k-labeled events that occurred in April–May 2016 (after the Kumamoto earthquake) and in late November 2016 (after the Fukushima-oki earthquake) have capital-letter flags. These account for only 0.2% of all k-labeled events. Based on these definitions and their actual usage, it is practical to select data using the K, k, and A labels instead of the data information flags. However, in this study, we followed the original labeling policy: data with K label were used only when accompanied by capital-letter data information flags, and data with the k or A labels only when accompanied by lowercase flags, even when not explicitly mentioned in this paper. Hereafter, we refer to these data as follows: arrival-time readings of events associated with K labels (with capital-letter flags) are termed “fully reviewed picks”; those associated with k labels (with lowercase flags) are termed “briefly reviewed picks”; and those associated with A labels are termed “automatically read picks.” Because the proportion of records with exceptional flags is extremely small, this treatment in training data selection is unlikely to affect the resultant model performance. The cumulative numbers of arrival-time readings and corresponding waveforms extracted under these conditions for events labeled K, k, and A are shown in Figure S1 , and the corresponding hypocenter distributions are shown in Figure S2 . As described in Naoi et al. ( 2024 ), the JMA catalog also includes events other than ordinary earthquakes, such as low-frequency earthquakes, which can be identified through the “subsidiary information” label assigned to each event. In this study, we extracted and analyzed only events labeled as ordinary earthquakes, whereas Naoi et al. ( 2024 ) used events of all subsidiary information types. 2.2 Data selection for model development In Section 3 , we modify the model architecture of the neural phase picker to achieve higher expressiveness and improve parameter efficiency. To enable repeated training and testing while modifying the model architecture, we prepared a reduced-size data set with a limited number of samples to shorten the computation time required for analysis. In this section, we describe the procedure used to construct this data set. The JMA unified data set consists of records from multiple observation networks, including ocean-bottom and broadband networks. As shown by Naoi et al. ( 2024 ), models trained on these data exhibit network-dependent performance. Therefore, when the proportion of data from each network varies within the data set, it becomes difficult to evaluate model performance consistently. To avoid this issue, we used only data from Hi-net, which provides long-term, stable, and densely distributed observations, to examine the model architecture. The data used for model development were extracted according to the following three criteria: (1) records from Hi-net stations; (2) events labeled K; and (3) availability of three-component waveform records. To suppress the influence of temporal changes in Japan’s routine network and the uneven distribution of large earthquakes and their aftershocks, the data were selected as follows: training data were extracted from 2002, 2005, 2008, 2011, 2014, 2017, 2020, and 2023; validation data from 2003, 2006, 2009, 2012, 2015, 2018, and 2021; and test data from 2004, 2007, 2010, 2013, 2016, 2019, and 2022. We avoided random extraction across the entire period because aftershock sequences, which often contain numerous events occurring within short intervals and confined regions, may produce highly similar waveforms. Such data could cause significant information leakage among the training, validation, and test data sets. In addition, to mitigate the imbalance caused by the Gutenberg–Richter law (Gutenberg and Richter 1944 ), in which the proportion of small-magnitude events becomes extremely large, we limited the number of samples extracted for each 0.1-magnitude bin to a maximum of 2000 for the training and test data sets and 400 for the validation data set. After applying these conditions, the final data sets contained 123,467 training samples, 26,939 validation samples, and 117,133 test samples. The FMD of these data is shown in Fig. 2 . Between magnitudes − 0.5 and 5.0, the number of waveforms is approximately uniform across magnitudes, ensuring a well-balanced data set. 2.3 Data selection to investigate the influence of using picks not subjected to full review as training data In Section 4.2 , we evaluate the influence of using the k and A hypocenter determination flags—introduced after the implementation of the PF method—for model training. To examine this effect, models trained with data sets compiled after the introduction of the PF method were evaluated using test data from before its implementation. We prepared the training data set from events in January–December 2019, the validation data set from October–December 2018, and the test data set from January–December 2015. From the data obtained in each of these periods, we constructed three data sets labeled K, k, and A, respectively, by extracting records that satisfied the following conditions: (1) records from Hi-net stations and (2) availability of three-component waveform records. The number of samples extracted under each condition is summarized in Table 2 , and their magnitude distributions are shown in Fig. 3 . The test data include records from reliably K-labeled events down to M ~ − 1, whereas in the training and validation data sets, records from K-labeled events are largely absent for small magnitudes and are almost entirely replaced by those from k- and A-labeled events. Table 2 Breakdown of the training, validation, and test data sets used to investigate the influence of including k- and A-labeled data for training (Section 4.2 ). Hypocenter determination flag Training Validation Test K 190,506 54,154 761,971 k 117,796 26,080 0 A 655,622 188,481 0 3 Modification of the architecture of the neural phase picker 3.1 Features of the PhaseNet architecture In this section, we progressively refine the original PhaseNet architecture (268,443 trainable parameters; Figure S3a) to develop a more parameter-efficient and scalable model that can be extended to larger-scale architectures, thereby achieving greater expressiveness. In this context, expressiveness corresponds to how effectively and how quickly the model can minimize the loss function during training. A typical approach to enhancing the expressiveness of deep learning models is to increase the number of convolutional layers or the number of filters (channels), as demonstrated by Naoi et al. ( 2024 ), who developed the PhaseNetWC model (1,070,899 trainable parameters). In the field of computer vision, however, various techniques have been developed to construct models with comparable expressiveness while reducing the number of trainable parameters, thereby improving parameter efficiency. Moreover, simply stacking additional layers can often impede training progress due to the vanishing-gradient problem, but techniques that mitigate vanishing gradients even in very deep networks are now well established (He et al. 2016 ). In this study, we incorporate these techniques into the model architecture. As pointed out by Zhu and Beroza ( 2019 ), the phase-picking problem can be regarded as a one-dimensional (1D) semantic segmentation task. They developed PhaseNet, which adapts U-Net (Ronneberger et al. 2015 )—a well-established architecture for semantic segmentation—to seismic waveform data. PhaseNet has been shown to achieve competitive or superior performance compared with other deep-learning-based phase pickers (Münchmeyer et al. 2022 ; García et al. 2022 ) and has been widely used to construct high-resolution seismic event catalogs (e.g., Kato 2024 ; Tan et al. 2021 ). As shown in Figure S3(a), PhaseNet consists of an encoder that compresses the time-series samples through 1D convolutions with a stride of 4 while increasing the number of channels, and a decoder that restores the time resolution using transposed 1D convolutions. The encoder and decoder are connected by skip connections, allowing high-resolution features lost during encoding to be recovered during decoding. PhaseNet has several distinctive features. The encoder exhibits an exceptionally high compression ratio: whereas most encoder–decoder architectures for image segmentation compress feature maps to between 1/8 and 1/32 of the input size (e.g., Ronneberger et al. 2015 ; Chen et al. 2018 ), PhaseNet reduces them to 1/256 of the original length. The 3001-sample input waveform used by Zhu and Beroza ( 2019 ) was compressed to only 11 samples. Such extreme compression helps capture long-range temporal dependencies but likely degrades the temporal resolution required for precise picking, an effect that is mitigated by the inclusion of skip connections. Another characteristic is that each convolutional layer employs filters with a large kernel size of 7, which likely also helps capture long-range dependencies. In conventional architectures, large receptive fields are typically achieved more efficiently by stacking several smaller convolutions (e.g., kernel size 3), making this design choice distinctive. Nevertheless, the recent success of Vision Transformers (ViT) in computer vision has renewed interest in architectures with large kernels (Ding et al. 2022 ; Liu et al. 2022a , b ). In this study, we take these characteristics of PhaseNet into account when refining the model architecture. Notably, this study focuses on convolutional neural network (CNN) based architectures derived from PhaseNet, although Vision Transformer–based approaches have recently achieved remarkable success in semantic segmentation tasks. For ViT-based models, we compare our results with SegPhase, a recently proposed ViT-based phase picker that has been shown to outperform PhaseNet (Katoh et al. 2025 ). 3.2 Development of baseline model The architectures of all models developed in this study are summarized in Table 3 . As a first step, we modified the PhaseNet architecture (Model 1) into a more manageable form. To simplify sample compression by strided convolution and the concatenation of feature maps, we assumed an input length of 4096 samples, which is a power of two. In the decoder, the transposed convolution layers were replaced with a simpler structure consisting of nearest-neighbor upsampling followed by 1D convolution. Each convolution was zero-padded so that, when no stride was applied, the number of samples remained unchanged before and after convolution. For clarity, layers producing outputs with the same number of samples were defined as belonging to the same “level,” and the number of channels within each level was kept constant, except where skip connections were applied. The model constructed with these modifications, shown in Fig. 4 (a), is referred to as the base model (Model 2) in this study. Table 3 Architectures of the neural phase pickers developed in this study. Model No. description No. of blocks No. of channels No. of training parameters 1 Original PhaseNet *Removed the unnecessary bias term in the first conv. layer. 268,435 2 Base model (simplified PhaseNet with stride = 4) [2, 2, 2, 2, 2] [8, 16, 32, 64, 128] 382,915 3 Increased conv. channels of Model 2. [2, 2, 2, 2, 2] [16, 32, 64, 128, 256] 1,528,323 4 Increased no. of conv. layers of Model 2 [4, 4, 4, 4, 4] [8, 16, 32, 64, 128] 689,443 5 Increased no. of conv. layers of Model 2 [6, 6, 6, 6, 6] [8, 16, 32, 64, 128] 995,971 6 Simplified PhaseNet (stride = 2) [2, 2, 2, 2, 2] [8, 16, 32, 64, 128] 382,915 7 Simplified PhaseNet (stride = 2) + ASPP module (inspired by DeepLab v3+), based on Model 6 [2, 2, 2, 2, 2] [8, 16, 32, 64, 128] 309,123 8 Replaced conv. layers of Model 7 with dw-conv. [2, 2, 2, 2, 2] [8, 16, 32, 64, 128] 67,872 9 Replaced conv. layers of Model 8 with Inverted Residual Bottleneck (expansion ratio = 4) [2, 2, 2, 2, 2] [8, 16, 32, 64, 128] 300,771 10 Replaced conv. layers of Model 8 with ConvNeXt Blocks (expansion ratio = 4) [2, 2, 2, 2, 2] [8, 16, 32, 64, 128] 288,032 11 Added skip connection to the Level 2 layer of Model 10 [2, 2, 2, 2, 2] [8, 16, 32, 64, 128] 288,400 12 Added skip connection to the Level 1 layer of Model 10 [2, 2, 2, 2, 2] [8, 16, 32, 64, 128] 288,152 13 Increased no. of conv. layers of Model 10 [2, 2, 4, 6, 4] [8, 16, 32, 64, 128] 704,544 14 Increased no. of conv. layers of Model 10 [2, 3, 3, 9, 3] [8, 16, 32, 64, 128] 666,208 15 Increased no. of conv. layers of Model 10 [2, 3, 3, 27, 3] [8, 16, 32, 64, 128] 1,269,856 16 Increased conv. channels of Model 13 [2, 2, 4, 6, 4] [12, 24, 48, 96, 192] 1,574,268 17 Increased conv. channels of Model 13 [2, 2, 4, 6, 4] [16, 32, 64, 128, 256] 2,788,984 18 Increased conv. channels of Model 14 [2, 3, 3, 9, 3] [12, 24, 48, 96, 192] 1,487,580 19 Increased conv. channels of Model 15 [2, 3, 3, 27, 3] [12, 24, 48, 96, 192] 2,835,420 20 Increased conv. channels of Model 19 [2, 3, 3, 27, 3] [16, 32, 64, 128, 256] 5,021,432 In the original PhaseNet (Model 1), the number of channels doubles in the first convolution of each level (Figure S3a). In the base model (Model 2), however, channel expansion is performed simultaneously with sample compression by stride, resulting in an approximately 1.5-fold increase in trainable parameters (382,915 in total). Figure 5 (a) shows the training losses over epochs for Model 1 and Model 2, trained using the data set prepared in Section 2.2 , with waveforms segmented into 4096-sample windows. To prevent the model from memorizing the arrival-time position, each waveform was extracted so that the P-wave arrival could appear anywhere within the entire window, determined using a uniform random offset. The inclusion of the S wave within the window was not strictly necessary. The ground-truth labels were assigned in the same manner as in Naoi et al. ( 2024 ), using the JMA arrival-time readings as peak positions with a height of 1.0 and representing them by Gaussian functions with a standard deviation of 0.1 s. Each window contained one pair of P and S phases unless the S wave extended beyond the window. This window-positioning and labeling procedure was also applied in subsequent analyses unless otherwise noted. As shown in Fig. 5 (a), the increased number of trainable parameters in Model 2 resulted in slightly lower loss values than those of Model 1. Because the purpose of this section is to examine the expressiveness of the model architecture, Fig. 5 presents only the training loss. Although not shown here, the validation loss began to diverge from the training loss at around 80 epochs for Model 1 and at around 60 epochs for Model 2, suggesting the onset of overfitting. As straightforward extensions to enhance the expressiveness of the base model, we constructed Model 3 by doubling the number of channels in the convolutional layers of Model 2 (1,528,323 trainable parameters), Model 4 by increasing the number of convolutional layers in each encoder level to four (689,443 parameters), and Model 5 by increasing the number to six (995,971 parameters). All other hyperparameters were identical to those of Model 2. The learning curves of these models, shown in Fig. 5 (a), indicate that Model 3—the one with the largest number of trainable parameters—showed the greatest loss reduction; however, compared with the models described later, the reduction in loss relative to the increase in parameter count was less efficient. Improving the efficiency of loss reduction is the focus of the following sections. 3.3 Adjustment of the sample-size compression ratio and modification for capturing long-range features As mentioned earlier, both the original PhaseNet (Model 1) and the base model (Model 2) compress seismic waveforms to 1/256 of the input length. For an input of 3001 samples, the feature maps are reduced to 11 samples, and for 4096 samples, to 16 samples. Because this high compression ratio likely sacrifices the resolution required for accurate picking, we reduced the ratio by changing the stride width of each layer from 4 to 2, resulting in a final compression ratio of 1/16. Figure 5 (b) shows the learning curve obtained for Model 6, in which the stride width was changed to 2 (Figure S3b). Compared with the base model (Model 2), the loss decreased more slowly, suggesting that the model’s expressiveness was reduced owing to the limited ability to capture long-range features. This ability to capture long-range dependencies is considered one of the key factors underlying the strong performance of ViTs in segmentation tasks (Khan et al. 2022 ) and in phase-picking problems (Katoh et al. 2025 ). Several studies have attempted to introduce the capability to learn long-range characteristics into convolution-based models for segmentation problems. One such approach is dilated convolution (Yu and Koltun 2016 ; Yu et al. 2017 ; Fig. 6 a), in which convolutional operations are applied at fixed intervals (rates) to expand the receptive field. Before the introduction of ViTs, the DeepLab v3 + model (Chen et al. 2018 ) gained popularity for segmentation tasks by introducing the Atrous Spatial Pyramid Pooling (ASPP) module (Fig. 6 b), which combines the outputs of multiple dilated convolutions performed with different rates to capture features at various spatial scales. In this study, we incorporated the ASPP module and the decoder of DeepLab v3 + to capture long-range information while maintaining stride-2 convolutions. The resulting model architecture (Model 7; 309,123 parameters) is shown in Fig. 4 (b), and its learning curve is superimposed in Fig. 5 (b). Despite having approximately 20% fewer parameters than Model 2, the ASPP-equipped model achieved lower loss values than Model 2 and also than Models 4 and 5, which have a greater number of layers (Fig. 5 b). 3.4 Parameter reduction using depthwise separable convolution Next, to reduce the number of trainable parameters, we introduced depthwise separable convolution (Sifre 2014 ; Chollet 2017 ; hereinafter referred to as dw-convolution). In a standard convolution, filtering is performed with kernels that have trainable parameters across both the spatial and channel dimensions (Fig. 7 a). In contrast, dw-convolution first applies a depthwise convolution, in which independent convolutions are performed for each channel, followed by a pointwise convolution using 1×1 kernels to combine information across channels (Fig. 7 b). This structure enables a substantial reduction in the number of parameters while maintaining the same input–output dimensions. We constructed Model 8 by replacing the convolutional layers of Model 7 with dw-convolution (Figure S3c). Only the convolution used in the global average pooling (GAP) within the ASPP module, the final convolution in the ASPP module, and the last convolutional layer of the network remained standard convolutions. As a result, the total number of parameters decreased to 67,872, but the training loss increased significantly, indicating reduced expressiveness (Fig. 5 c). This reduction in expressiveness is addressed in the next section by increasing the number of layers based on the inverted residual bottleneck (Sandler et al. 2018 ) and the ConvNeXt block (Liu et al. 2022b ). 3.5 Increasing model scalability by introducing inverted residual bottlenecks and ConvNeXt blocks with residual connections As described in Section 3.1 , simply increasing the number of convolutional layers does not always lead to better performance and can often cause degradation. If additional layers could perfectly behave as an identity mapping, such degradation would not occur; therefore, an important limitation of CNNs is that their layers cannot easily learn identity mappings. To address this issue, He et al. ( 2016 ) introduced residual connections and developed the ResNet model, which enables stable training of much deeper networks. In this study, we also constructed scalable models that can stack more layers by incorporating residual connections. Here, we enhance expressiveness by stacking inverted residual bottlenecks (Fig. 8 a) and ConvNeXt blocks (Fig. 8 b) with residual connections. The inverted residual bottleneck first expands the number of channels in the feature maps using 1×1 (pointwise) convolutions, applies a depthwise convolution with multiple channels, and then performs another 1×1 convolution to produce output feature maps with the desired number of channels. By placing pointwise convolutions before and after the depthwise convolution, this block achieves high expressiveness while keeping the number of parameters moderate, because the channel dimensions outside the block remain compact. We constructed Model 9 (Fig. 4 c) using inverted residual bottlenecks that have the same input and output feature-map sizes as the convolutional layers in the encoder parts of Models 7 and 8. As shown in Fig. 8 (a), the channel expansion ratio in the intermediate layer was set to 4, and the kernel size to 7. Model 9 achieved loss values comparable to those of Model 7 (Fig. 5 c), with slightly fewer trainable parameters. Next, we constructed Model 10, in which each inverted residual bottleneck was replaced with a ConvNeXt block. The ConvNeXt block reverses the order of the first and second layers in the inverted residual bottleneck and modifies the activation functions (Fig. 8 b). These architectural refinements incorporate design insights gained from ViTs, allowing convolution-based networks to achieve performance comparable to ViTs (Liu et al. 2022b ). The total number of parameters of Model 10 was 4.2% smaller than that of Model 9 (Table 3 ), while achieving a lower training loss (Fig. 5 c). In the following sections, we use Model 10 as the base for further improvement. We refer to this type of architecture as PhaseNeXt, and to Model 10 specifically as PhaseNeXt-S. 3.6 Skip connection In all models developed so far, only the Level-3 skip connection was passed to the decoder, following the structure of DeepLab v3+. To examine the effect of skip connections from shallower layers, we constructed Model 11, which added a Level-2 skip connection, and Model 12, which added a Level-1 skip connection, to Model 10. Training results showed no noticeable difference in the learning curves (Fig. 5 c) compared with Model 10. Therefore, in this study, skip connections from Levels 1 and 2 were not included. It should also be noted that, whereas the original PhaseNet used stride-4 convolutions, the present models employ stride-2 convolutions. Consequently, the two-step upsampling from Level 3 in Models 6–12 corresponds to only one step of resolution recovery in the original PhaseNet. This may be the reason why the additional skip connections did not make a significant difference. 3.7 Scaling model size by stacking blocks and widening channels Since the proposal of ResNet (He et al. 2016 ), it has become common to enhance model expressiveness by stacking multiple standardized blocks, such as inverted residual bottlenecks and ConvNeXt blocks. In such architectures, expressiveness can be easily scaled by adjusting the number of blocks and channels. Following common design practices observed in modern convolutional networks (He et al. 2016 ; Sandler et al. 2018 ), fewer blocks are typically used in shallower levels—where feature maps are larger—and more blocks are stacked in deeper levels to balance representational power and computational efficiency. Furthermore, as reported by Tan and Le ( 2019 ), simultaneously scaling both the depth (number of layers) and the width (number of channels) provides a more efficient way to enhance expressiveness than increasing either the depth or the width alone. Based on Model 10, we developed and trained a series of larger models (Models 13–19; Table 3 ) by increasing the number of layers (blocks), the number of channels, or both. The corresponding learning curves are shown in Fig. 5 (d). As the model size increased, the loss decreased, and the minimum loss values were achieved by Models 17 and 19, which contained the largest number of parameters. As demonstrated by Models 3–5, such low loss values are difficult to achieve simply by increasing the number of channels or layers in the base model, highlighting the importance of the architectural refinements. 4 Data selection and preprocessing schemes 4.1 Preprocessing of signal amplitudes When inputting waveforms into neural phase pickers, normalization or standardization is often applied to each channel individually (Zhu and Beroza 2019 ). Such preprocessing helps stabilize and accelerate model training by aligning the scales of input values. However, in seismic waveforms, the amplitude ratio among channels carries important information. For example, the initial motion of P wave is aligned with the ray path, and this linearity may provide a key indicator for identifying the P phase. Therefore, applying normalization that preserves the inter-channel amplitude ratio may improve the performance of phase-picking models. In addition, Naoi et al. ( 2024 ) reported that the picking performance of PhaseNet and PhaseNetWC decreased as event magnitude increased. Although they attributed this magnitude dependence to the shortage of large-magnitude events in the training data—a result of the Gutenberg–Richter law—we consider another possible factor: amplitude normalization applied during preprocessing may reduce P-wave picking performance. Specifically, when normalized waveforms from large-amplitude events are fed into the model, the relative change from noise to the P-wave onset becomes smaller within the overall amplitude range, potentially leading to degraded P-wave picking performance for high-magnitude events. To mitigate this issue, normalization based on the noise level with amplitude saturation at a certain threshold or a soft-clipping technique (Zhu et al. 2019 ) could be applied; however, these approaches would inevitably lead to some loss of waveform information. Instead, in this study, we examined whether this degradation in performance could be mitigated by simultaneously inputting six channels—three fully normalized components and three additional components in which the P-wave portion was amplified using noise-level-based normalization with amplitude clipping. Based on these ideas, we trained the same model under four input configurations and compared their performance: (1) three-component waveforms standardized independently for each channel (z-score normalization); (2) three-component waveforms standardized while preserving inter-channel amplitude ratios (each channel was mean-corrected, and the three components were jointly normalized by their overall standard deviation); (3) six-component waveforms composed of the three individually standardized components and an additional three P-wave-amplified components; and (4) six-component waveforms composed of the three components standardized while preserving inter-channel amplitude ratios and another three P-wave-amplified components that also preserve inter-channel amplitude ratios. The approaches used by Zhu and Beroza ( 2019 ) and Naoi et al. ( 2024 ) correspond to case (1). The amplification of waveforms was implemented by standardizing the first 150 samples of each waveform segment using their own mean and standard deviation. As described in Section 3.2 , the P-wave arrival was generally positioned at a random location within the whole of the cutting window. However, in this analysis, we prepared the data so that the P-wave arrival did not occur within the first 150 samples. This analysis was conducted using the data set prepared in Section 2.2 , with the Model 10 architecture trained under the four input configurations described above. In cases (3) and (4), the number of input channels was expanded from three to six in the model illustrated in Fig. 4 (c). The models that achieved the minimum validation loss in each training run were used for performance evaluation on the test data. As shown in Fig. 2 , the number of samples in the data set exhibited no magnitude dependence within the range of approximately − 0.5 < M < 5, allowing performance evaluation without bias arising from the training-data imbalance noted by Naoi et al. ( 2024 ). The evaluation followed the approach of Naoi et al. ( 2024 ), in which true positives, false positives, and false negatives were defined according to Sun et al. ( 2023 ) (Table S1 ), and performance was assessed using precision–recall (PR) curves. Figure S4 shows the PR curves for the models trained using the four input configurations. Except for the P-wave results at M < 0 (whose lower performance likely reflects the lower signal-to-noise ratio of the waveforms, as noted by Naoi et al. 2024 ), a magnitude-dependent trend similar to that observed by Naoi et al. ( 2024 ) remained evident for both P and S phases, even though the training data set was designed to minimize magnitude dependence. This tendency did not improve even when the models were trained using amplified waveforms (cases 3 and 4) or when inter-channel amplitude ratios were preserved (cases 2 and 4). These results indicate that the degradation of picking performance with increasing magnitude is not caused by deficiencies in the preprocessing or composition of the training data, but rather reflects an intrinsic property of the waveforms themselves. Because the differences in performance were minor, subsequent analyses were conducted using configuration (1)—the simplest approach, consistent with that adopted by Zhu and Beroza ( 2019 ) and Naoi et al. ( 2024 ). 4.2 Incorporating briefly reviewed and automatically picked data into training As described in Section 2.1 , picks associated with K-labeled events (i.e., events subjected to full manual review) were lacking for small-magnitude events after the introduction of the PF method, likely resulting in reduced performance for small events recorded by the networks developed thereafter. To examine whether this performance degradation could be mitigated by incorporating briefly reviewed or automatically processed picks into the training data set, we trained Model 10 using three data sets derived from the Hi-net data prepared in Section 2.3: (1) data associated with K-labeled events, (2) data associated with either K- or k-labeled events, and (3) data associated with K-, k-, or A-labeled events. Figure 9 shows the PR curves obtained using Model 10 at the epoch that achieved the minimum validation loss in the training with each of the three data sets. The test data consisted of events from 2015, all labeled as K. The model trained with data set (1) exhibited lower detection performance for small events, particularly those with M < 0, reflecting the shortage of such small-magnitude events in its training data. In contrast, training with data sets (2) and (3) substantially improved detection performance for these small events. These results indicate that combining briefly reviewed and automatically read picks in the training process can compensate for the lack of K-labeled data in recently deployed observation networks. Although data set (3) contained more training samples than data set (2), its performance deteriorated, particularly for very small events. In the following analyses, we use the data set combining K and k labels for model training. 5 Training of modified models using a large JMA unified data set 5.1 Learning curves Based on the findings obtained in Sections 3 and 4 , we trained neural phase pickers using arrival-time records listed in the JMA unified catalog for 2002–2012 and 2014–2023, along with the corresponding waveforms extracted from continuous records. For validation and testing, we used the 2013 data set, in which all picks were fully reviewed: data from January to June were used for validation, and those from July to December for testing. As described in the previous section, events with hypocenter determination flags of K and k were used for training. The total number of waveforms used for training was 24,469,146 (22,610,301 with K flags and 1,858,845 with k flags), approximately four times larger than that used by Naoi et al. ( 2024 ), with 820,328 waveforms for validation and 1,229,716 for testing (all with K flags). The 2013 data were also used as test data in Naoi et al. ( 2024 ) and were not included in their training. Therefore, using the same test data set, we compared the performance of the models trained in this study with those trained by Naoi et al. ( 2024 ). We refer to the PhaseNet retrained by Naoi et al. ( 2024 ) using the eight-year JMA data set as PhaseNet-J, and to the PhaseNetWC trained by them as PhaseNetWC-J. Using this data set, we trained three models for 60 epochs under the same computational environment and training parameters: the base model (Model 2), PhaseNeXt-S (Model 10), and PhaseNeXt-M (Model 19). PhaseNeXt-S has about 75% of the trainable parameters of the base model, whereas PhaseNeXt-M has approximately 7.4 times more. Training was conducted using two NVIDIA A40 GPUs. With a batch size of 1024, the training times were 119 h for the base model, 126 h for PhaseNeXt-S, and 398 h for PhaseNeXt-M. For comparison, we also trained SegPhase, a ViT-based model proposed by Katoh et al. ( 2025 ), and Model 20, in which the number of channels in PhaseNeXt-M was further increased. For SegPhase, we adopted the convolutional stride settings specified by Katoh et al. ( 2025 ; their Fig. 2 ), with values of (st₁, st₂, st₃) = (2, 2, 2), and the total number of trainable parameters in this configuration was 333,107. Because of their substantial computational demands, these two models could not be trained under the same environment or batch-size settings as the previous three models. SegPhase was trained for 20 epochs on the same system with a reduced batch size of 512, which took 300 h; extrapolation suggests that approximately 648 h would have been required to train it for 60 epochs. Model 20 was trained on the JMA supercomputer equipped with eight NVIDIA A100 GPUs, using a batch size of 4096 for 20 epochs. The resulting learning curves are shown in Fig. 10 . As expected from the training-loss differences shown in Fig. 5 , both the training and validation losses decreased successively from the base model to PhaseNeXt-S and PhaseNeXt-M. Notably, in all cases, the validation loss was smaller than the training loss—opposite to the trend typically observed in deep-learning training. This suggests that the validation data were easier for the model to pick than the training data. This tendency is likely because the validation data consisted exclusively of K-labeled events from 2013. Even for PhaseNeXt-M, no signs of overfitting were observed, similar to the findings reported by Naoi et al. ( 2024 ) for PhaseNet and PhaseNetWC, suggesting that further scaling of the model size may lead to improved performance. Indeed, in training of the larger model (Model 20), the minimum validation loss was slightly smaller than that of PhaseNeXt-M; however, such large models require substantially longer training and prediction times. Although further enlarging the model may yield additional performance gains, practical applications would require model pruning or other post-training compression techniques to reduce computational costs. Regarding SegPhase, although it has a comparable number of trainable parameters to the base model, its training cost was even higher than that of PhaseNeXt-M due to the inclusion of multi-head attention, whereas its performance only slightly exceeded that of PhaseNeXt-S and remained lower than that of PhaseNeXt-M. While ViT-based architectures may demonstrate superior performance to CNN-based models when trained on extremely large data sets with sufficient computational resources—as achieved in segmentation tasks and large language models—the advantage of ViT-based approaches was not evident under the data volume and computational environment considered in this study. 5.2 PR curves for the 2013 test data set For the base model, PhaseNeXt-S, and PhaseNeXt-M, we adopted the results corresponding to the minimum validation losses (epoch 20 for the base model and epoch 53 for both PhaseNeXt-S and PhaseNeXt-M) in the subsequent analysis. The resultant models, along with the PhaseNetWC-J model developed by Naoi et al. ( 2024 ), were evaluated using the test data from July–December 2013. The PR curves for the test data set are shown in Fig. 11 . The performances of the base model and PhaseNetWC-J were nearly identical, whereas clear improvements—indicated by PR curves shifting toward the upper-right corner—were observed for PhaseNeXt-S and PhaseNeXt-M. Models that achieved lower loss values in the training curves shown in Fig. 5 tended to exhibit better performance. It should be noted that the 2013 test data consisted of records acquired before the deployment of DONET2 (Dense Ocean floor Network for Earthquakes and Tsunamis; NIED 2019c) and S-net, for which the influence of incorporating k-labeled events is expected to be more pronounced. 5.3 Comparison of application results for the JMA unified data set Naoi et al. ( 2024 ) compared the hypocenter determination results obtained using the JMA automatic processing routine with those from a workflow in which the picking method was replaced by their retrained PhaseNet-J, based on data recorded on March 31, 2023, after adjusting the P- and S-phase thresholds to yield approximately the same number of picks. In this study, we applied the same approach to PhaseNetWC-J, PhaseNeXt-S, and PhaseNeXt-M to examine whether the present models outperform PhaseNetWC-J. The numbers of picks and determined hypocenters are summarized in Table 4 , and the resulting hypocenter distributions are shown in Fig. 12 . The “Auto-PF” result shown in the table and figure refers to the catalog developed using the JMA-adopted automatic processing algorithm. Although the PhaseNet-J results shown for comparison differ slightly from those reported by Naoi et al. ( 2024 ), this discrepancy arises because Naoi et al. ( 2024 ) applied an additional process to read amplitudes and repick arrival times using the same method employed in the development of the JMA unified catalog (this repicking process slightly adjusted the PhaseNet-J arrival times, and when repicking failed, the original PhaseNet-J picks were retained). Because this study did not apply such post-processing, the comparison is more direct; however, this difference has only a minor influence on the overall discussion. PhaseNeXt-S determined nearly the same number of events as PhaseNetWC-J, whereas PhaseNeXt-M achieved the largest number of determined hypocenters—approximately 3.5 times as many as those listed in the JMA unified catalog. Table 4 Numbers of picks and cataloged events obtained from continuous data recorded on December 31, 2023 (updated from Table 4 in Naoi et al. 2024 ). p th for P-wave p th for S-wave No. of P-picks No. of S-picks No. of hypocenters Matched Unified catalog - - - - 636 - Auto-PF - - 427,436 735,105 911 521(2) PhaseNet-PF 0.250 0.106 436,107 735,539 1747 561(3) PhaseNetWC-PF 0.310 0.090 432,814 736,685 1909 550(4) PhaseNeXt-S (Model 10)-PF 0.234 0.031 427,283 729,928 1878 552(4) PhaseNeXt-M (Model 19)-PF 0.200 0.018 430,500 758,514 2135 564(4) * ‘Matched' indicates that the corresponding events were found in the JMA unified catalog (Naoi et al. 2024 ). The numbers shown in brackets represent duplicate counts (i.e., cases where multiple detected events are associated with a single event in the unified catalog). In the above analysis, the PhaseNeXt models required significantly lower peak thresholds compared with PhaseNetWC-J, especially for the S wave, to obtain approximately the same number of picks. To evaluate how these picks with low peak values contributed to hypocenter determination, Figs. 13 (a) and 13(b) show the cumulative number of effective picks—defined as those actually used for hypocenter determination—whereas Figs. 13 (c) and 13(d) present the proportion of effective picks, closely corresponding to the “utilization rate” defined by Katoh et al. ( 2025 ), for each peak-value bin with an interval of 0.1. From Figs. 13 (a) and 13(b), it is evident that the number of effective picks obtained with the PhaseNeXt models increased relative to those obtained with PhaseNet-J and PhaseNetWC-J. For PhaseNeXt-M, the advantage was found for peak values below 0.8 for the P phase and between 0.2 and 0.8 for the S phase. For PhaseNeXt-S, a similar number of effective picks to those obtained with PhaseNeXt-M was found in the same range for the S wave. Figure 13 (b) indicates that S-phase picks with very low peak values below 0.1 significantly contribute to the large number of effective picks obtained with the PhaseNeXt models. When the same threshold as that applied to PhaseNetWC-J was used for PhaseNeXt-M, the number of available picks decreased to approximately 42% (269,317 for P waves and 231,342 for S waves, totaling 500,659 picks). Nevertheless, the number of determined hypocenters remained identical to that obtained with PhaseNetWC-J (1909 events). For PhaseNeXt-S, using the same threshold resulted in 92% of the hypocenters (1751 events) being determined from 54% of the picks (308,545 for P waves and 321,507 for S waves, totaling 630,052 picks). These results clearly demonstrate the higher precision of the PhaseNeXt models compared with PhaseNetWC-J. Figures 13 (c) and 13(d), which show the percentage of effective picks, directly demonstrate the high precision of the PhaseNeXt models. This advantage is particularly evident for the S phase, where PhaseNeXt-S and PhaseNeXt-M clearly outperform PhaseNet-J and PhaseNetWC-J. Although this characteristic is favorable for earthquake catalog compilation, in networks with many stations and highly accurate phase association, even a large number of false picks has only a minor effect on the final number of determined hypocenters. This likely explains the relatively small difference in the number of determined events among PhaseNetWC-J, PhaseNeXt-S, and PhaseNeXt-M shown in Table 4 . The advantage of the PhaseNeXt models is expected to become more pronounced in cases where the number of stations is limited and false detections have a greater impact. In any case, PhaseNeXt-M exhibited the best performance among all models compared. However, owing to its larger architecture, additional computational cost is required not only for training but also for prediction. In our test applying the model to continuous waveform data, PhaseNeXt-M required approximately 4.3 times more computation time than PhaseNetWC-J. This additional cost in picking may become a constraint in near–real-time catalog development or large-scale applications, although the reduced number of false detections produced by PhaseNeXt-M helps to lower the computational cost of the subsequent phase association process. For prediction, PhaseNeXt-S required approximately 1.3 times more computation time than PhaseNetWC-J, even though PhaseNeXt-S has fewer parameters than PhaseNetWC-J. Its computational load appears to have increased due to architectural differences. As is often the case with large language models, practical applications may require selecting an appropriate model by considering the trade-off between computational cost and accuracy. 6 Discussion and conclusions In this study, we extended the work of Naoi et al. ( 2024 ) on training deep-learning-based phase pickers using the JMA unified data set by incorporating seismic records spanning 2002–2023. From more than 25 million waveform records, we developed and trained a new neural phase picker, PhaseNeXt. PhaseNeXt builds on PhaseNet but integrates recent advances in computer vision—specifically, the ConvNeXt block and the Atrous Spatial Pyramid Pooling (ASPP) module—to achieve higher expressiveness and greater parameter efficiency. During training, we demonstrated that including briefly reviewed picks, in addition to those subjected to detailed manual inspections, further improved performance. The final model, PhaseNeXt-M, outperformed the existing PhaseNet-J and PhaseNetWC-J models and, when combined with the PF method, detected approximately 3.5 times more hypocenters than those listed in the JMA unified earthquake catalog, while using a comparable number of picks to the JMA automatic processing. Although PhaseNeXt-M achieved high performance, several challenges remain. To further improve performance and facilitate practical implementation, the following directions can be considered: (1) enhancing data augmentation: training with techniques such as noise addition (Naoi et al. 2025 ) and waveform superposition (Kim et al. 2023 ) may improve detection performance under low signal-to-noise conditions or during periods of frequent seismicity; (2) mitigating label imbalance: phase picking is inherently a highly imbalanced classification problem, as most labels correspond to noise. As pointed out by Katoh et al. ( 2024 ), introducing weighted learning could improve both training efficiency and model performance; (3) model compression: applying techniques such as pruning and knowledge distillation may enable the development of lightweight models capable of real-time processing without significant loss of accuracy; and (4) improving performance for large events. Advancing these improvements and reconstructing earthquake catalogs from the extensive continuous waveform archives accumulated to date could further enhance seismic monitoring systems, and the integration of PhaseNeXt into such systems could lead to more accurate and comprehensive earthquake catalogs in the future. Abbreviations JMA Japan Meteorological Agency PF Phase combination Forward search CNN Convolutional neural network ViT Vision Transformer GPU Graphics Processing Unit ASPP Atrous Spatial Pyramid Pooling BN Batch Normalization dw-convolution Depthwise separable convolution GAP Global average pooling PR Precision–recall FMD Frequency magnitude distribution Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Availability of data and materials The base model, PhaseNeXt-S, and PhaseNeXt-M models trained using the 20-year JMA unified data set are available at https://github.com/mktnaoi/JMAuniPicker. Seismic waveform data related to the routine seismic observations in Japan (including a network other than Hi-net) are available at http://www.hinet.bosai.go.jp/?LANG=en. The JMA unified catalog is available at https://www.data.jma.go.jp/svd/eqev/data/bulletin/index.html. Competing interests The authors declare no competing interests. Funding This work was supported by JSPS KAKENHI (Grant Nos. JP21H01191, JP20K14565, JP24H01027, and JP25K01083) and by the Ministry of Education, Culture, Sports, Science, and Technology (MEXT) of Japan under its Earthquake and Volcano Hazards Observation and Research Program. Authors' contributions MN was the primary contributor to analysis and manuscript writing. KS and KT provided detailed information on handling the JMA unified data set and contributed to catalog development and figure preparation. All authors contributed to the writing and have read and approved the final manuscript. Acknowledgements This study used data deriving from the National Research Institute for Earth Science and Disaster Resilience (2019a–d), Hokkaido University, Hirosaki University, Tohoku University, the University of Tokyo, Nagoya University, Kyoto University, Kochi University, Kyushu University, Kagoshima University, National Institute of Advanced Industrial Science and Technology (AIST), Geospatial Information Authority of Japan, Japan Agency for Marine-Earth Science and Technology (JAMSTEC), Aomori Prefecture, Tokyo Metropolitan Government, Shizuoka Prefecture, Yokohama City (Kanagawa Prefecture), Hot Springs Research Institute of Kanagawa Prefecture, Association for the Development of Earthquake Prediction, Group for Urgent Joint Seismic Observation of the 2016 Kumamoto Earthquake, and Japan Meteorological Agency. Regarding the analysis, we employed a large continuous seismic data analysis system (Nakagawa et al. 2016) at the Earthquake Research Institute, University of Tokyo (ERI JURP 2024-F3-12). We used SeisBench (Woollam et al. 2022) based on the PyTorch library for deep-learning analysis. References Aoi S, Asano Y, Kunugi T et al (2020) MOWLAS: NIED observation network for earthquake, tsunami and volcano. Earth Planets Space 72:126. https://doi.org/10.1186/s40623-020-01250-x Chen LC, Zhu Y, Papandreou G et al (2018) Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European Conference on Computer Vision. Springer, pp 833–851. https://doi.org/10.1007/978-3-030-01234-2_49 Chollet F (2017) Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, pp 1800-1807. https://doi.org/10.1109/CVPR.2017.195 Ding X, Zhang X, Han J, Ding G (2022) Scaling up your kernels to 31×31: Revisiting large kernel design in CNNs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, pp 11963–11975. https://doi.org/10.1109/CVPR52688.2022.01166 García JE, Fernández-Prieto LM, Villaseñor A et al (2022) Performance of deep learning pickers in routine network processing applications. Seismol Res Lett 93:2529–2542. https://doi.org/10.1785/0220210323 Gutenberg B, Richter CF (1944) Frequency of earthquakes in California. Bull Seism Soc Am 34:185–188. https://doi.org/10.1785/BSSA0340040185 Hara S, Fukahata Y, Iio Y (2019) P-wave first-motion polarity determination of waveform data in western Japan using deep learning. Earth Planets Space 71:1–11. https://doi.org/10.1186/s40623-019-1111-x He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE. https://doi.org/10.1109/CVPR.2016.90 Japan Meteorological Agency (2025) The Seismological Bulletin of Japan. https://www.data.jma.go.jp/svd/eqev/data/bulletin/index_e.html. Accessed 1 Sep 2025 Kato A (2024) Implications of fault-valve behavior from immediate aftershocks following the 2023 Mj6.5 earthquake beneath the Noto Peninsula, central Japan. Geophys Res Lett 51:e2023GL106444. https://doi.org/10.1029/2023GL106444 Katoh S, Nagao H, Imaizumi M (2024) Towards addressing challenges in seismic wave arrival-time picking models using deep learning. In: Proceedings of the Annual Conference of the Japanese Society for Artificial Intelligence. https://doi.org/10.11517/pjsai.JSAI2024.0_3L1OS3a02 Katoh S, Iio Y, Nagao H et al (2025) SegPhase: development of arrival time picking models for Japan’s seismic network using the hierarchical vision transformer. Earth Planets Space 77:118. https://doi.org/10.1186/s40623-025-02249-y Khan S, Naseer M, Hayat M et al (2022) Transformers in vision: A survey. ACM Comput Surv 54:1–41. https://doi.org/10.1145/3505244 Kim A, Nakamura Y, Yukutake Y et al (2023) Development of a high-performance seismic phase picker using deep learning in the Hakone volcanic area. Earth Planets Space 75:85. https://doi.org/10.1186/s40623-023-01840-5 Kubo H, Naoi M, Kano M (2024) Recent advances in earthquake seismology using machine learning. Earth Planets Space 76:36. https://doi.org/10.1186/s40623-024-01982-0 Liu S, Chen T, Chen X et al (2022a) More ConvNets in the 2020s: Scaling up kernels beyond 51×51 using sparsity. https://arxiv.org/abs/2207.03620. [cs.CV] Liu Z, Mao H, Wu CY et al (2022b) A ConvNet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, pp 11966–11976. https://doi.org/10.1109/CVPR52688.2022.01167 Münchmeyer J, Woollam J, Rietbrock A et al (2022) Which picker fits my data? A quantitative evaluation of deep learning based seismic pickers. J Geophys Res. https://doi.org/10.1029/2021JB023499 Nakagawa S, Tsuruoka H, Kato A, Sakai S, Hirata N (2016) A petabyte-scale large continuous seismic data analyzing system. Bull Earthq Res Inst 91:1–9. https://doi.org/10.15083/0000032408 (in Japanese with English abstract) Naoi M, Tamaribuchi K, Shimojo K et al (2024) Neural phase picker trained on the Japan Meteorological Agency unified earthquake catalog. Earth Planets Space 76:150. https://doi.org/10.1186/s40623-024-02091-8 Naoi M, Hirano S, Chen Y (2025) High-resolution monitoring of hydraulically induced acoustic emission activities using neural phase picking and matched filter analysis. Prog Earth Planet Sci 12:24. https://doi.org/10.1186/s40645-025-00696-5 National Research Institute for Earth Science and Disaster Resilience (2019a) NIED Hi-net. https://doi.org/10.1759/NIED.0003 National Research Institute for Earth Science and Disaster Resilience (2019b) NIED S-net. https://doi.org/10.17598/NIED.0007 National Research Institute for Earth Science and Disaster Resilience (2019c) NIED DONET. https://doi.org/10.1759/NIED.0008 National Research Institute for Earth Science and Disaster Resilience (2019d) NIED F-net. https://doi.org/10.17598/NIED.0005 Ronneberger O, Fischer P, Brox T (2015) U-Net: Convolutional networks for biomedical image segmentation. In: Proceedings of the Medical Image Computing and Computer-Assisted Intervention. Springer, 9351:234–241. https://doi.org/10.1007/978-3-319-24574-4_28 Sandler M, Howard A, Zhu M et al (2018) MobileNetV2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, pp 4510–4520. https://doi.org/10.1109/CVPR.2018.00474 Sifre L (2014) Rigid-motion scattering for image classification. Ph.D. thesis, École Polytechnique, Palaiseau, France. Sun H, Ross ZE, Zhu W, Azizzadenesheli K (2023) Phase neural operator for multi-station picking of seismic arrivals. Geophys Res Lett 50:e2023GL106434. https://doi.org/10.1029/2023GL106434 Suzuki R, Uchida N, Zhu W et al (2025) The forearc seismic belt: A fluid pathway constraining down-dip megathrust earthquake rupture. Science 389:190–194. https://doi.org/10.1126/science.adt6389 Tamaribuchi K (2018) Evaluation of automatic hypocenter determination in the JMA unified catalog. Earth Planets Space 70:1–10. https://doi.org/10.1186/s40623-018-0915-4 Tamaribuchi K, Moriwaki K, Ueno H, Tsukada S (2016) Automatic hypocenter determination for the Seismological Bulletin of Japan using Bayesian estimation. Quart J Seis 79:1–13 Tan YJ, Waldhauser F, Ellsworth WL et al (2021) Machine-learning-based high-resolution earthquake catalog reveals how complex fault structures were activated during the 2016–2017 central Italy sequence. The Seismic Record 1:11–19. https://doi.org/10.1785/0320210001 Tan M, Le QV (2019) EfficientNet: Rethinking model scaling for convolutional neural network. In: Proceedings of the 36th International Conference on Machine Learning, pp 6105–6114 Woollam J, Münchmeyer J, Tilmann F et al (2022) SeisBench—a toolbox for machine learning in seismology. Seismol Res Lett 93:1695–1709. https://doi.org/10.1785/0220210324 Yu F, Koltun V (2016) Multi-scale context aggregation by dilated convolutions. In: Bengio Y, LeCun Y (eds) Proceedings of the 4th International Conference on Learning Representations, San Juan, Puerto Rico. Yu F, Koltun V, Funkhouser T (2017) Dilated residual networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, pp 636–644. https://doi.org/10.1109/CVPR.2017.75 Zhu W, Beroza GC (2019) PhaseNet: A deep-neural-network-based seismic arrival-time picking method. Geophys J Int 216:261–273. https://doi.org/10.1093/gji/ggy423 Zhu L, Peng Z, McClellan J et al (2019) Deep learning for seismic phase detection and picking in the aftershock zone of 2008 M7.9 Wenchuan earthquake. Phys Earth Planet Inter 293:106261. https://doi.org/10.1016/j.pepi.2019.05.004 Supplementary Files 251121Summlement.pdf FigGA.png Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Major Revision 23 Mar, 2026 Reviewers agreed at journal 18 Dec, 2025 Reviewers invited by journal 18 Dec, 2025 Editor assigned by journal 27 Nov, 2025 First submitted to journal 21 Nov, 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-8174647","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":562359434,"identity":"694edaca-8bd6-49b8-b348-592e9c256bca","order_by":0,"name":"Makoto Naoi","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA5ElEQVRIie2QMQuCQBiGvxB0OXE9qR9x4VBC1F8RDpx0bWkoCJxqb+iHNCof1BK6NhZBk4NtDQ6dBVJDmlvEPdMz3MPdewASyY8SZgAEKCgAqpACWpNEq8aJQp6nHkk9PQ0jHOTDDrTnaJGxEA2PYG8+J/bCddAPOIHOlnMSCyEuA3P/OWGhx9CfhWKLZ6EeFI/0AMygIklShv28TKYEjLQmOYhbQH0mXA+wkLrk4kRLsUWlLu+u452QCwsrtyQcs1s+HBmUI03Hk5Fh8NPZrPixEvVV0Jx9kbzRujZOJBKJ5I+5A3MORfOyI5f/AAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0001-9488-9266","institution":"Hokkaido University: Hokkaido Daigaku","correspondingAuthor":true,"prefix":"","firstName":"Makoto","middleName":"","lastName":"Naoi","suffix":""},{"id":562359435,"identity":"f991b61f-06d3-445f-b7e5-96d1d3faa686","order_by":1,"name":"Kengo Shimojo","email":"","orcid":"","institution":"Meteorological Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Kengo","middleName":"","lastName":"Shimojo","suffix":""},{"id":562359436,"identity":"1c5ce439-1a3d-4d85-81c5-9559ec5f64ca","order_by":2,"name":"Koji Tamaribuchi","email":"","orcid":"","institution":"Japan Meteorological Agency","correspondingAuthor":false,"prefix":"","firstName":"Koji","middleName":"","lastName":"Tamaribuchi","suffix":""}],"badges":[],"createdAt":"2025-11-21 14:39:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8174647/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8174647/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":98817692,"identity":"ba6aeaec-691a-43ed-ada2-cd2fc430fe71","added_by":"auto","created_at":"2025-12-22 16:39:09","extension":"xml","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10353,"visible":true,"origin":"","legend":"","description":"","filename":"epspEPSPD2500383.xml","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/bf057d72359f3f9c63bec3ca.xml"},{"id":98817681,"identity":"692bead3-eec0-4108-9847-100eef5626e9","added_by":"auto","created_at":"2025-12-22 16:39:08","extension":"xml","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1555,"visible":true,"origin":"","legend":"","description":"","filename":"EPSPD250038311596.go.xml","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/1f9d60aa1f18126c14bce283.xml"},{"id":98817685,"identity":"f991861e-c8e4-479a-9847-45f01e74e269","added_by":"auto","created_at":"2025-12-22 16:39:08","extension":"xml","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":809,"visible":true,"origin":"","legend":"","description":"","filename":"EPSPD2500383Import.xml","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/c7b6e802ee8bf865f2e5a280.xml"},{"id":99307230,"identity":"8dc28ff7-f755-4471-b11b-f2039dd1abda","added_by":"auto","created_at":"2025-12-31 16:05:50","extension":"xml","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":139722,"visible":true,"origin":"","legend":"","description":"","filename":"EPSPD25003830enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/118e4cbfcd3899dc39443db2.xml"},{"id":98817697,"identity":"8e104055-9259-4799-99f1-cdf1cfe4893f","added_by":"auto","created_at":"2025-12-22 16:39:09","extension":"pdf","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":461045,"visible":true,"origin":"","legend":"","description":"","filename":"Fig01JMAFMD.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/18de59af2ec565f11761cdc8.pdf"},{"id":99307740,"identity":"04dd622e-699e-4824-b460-0ec645978f08","added_by":"auto","created_at":"2025-12-31 16:06:41","extension":"pdf","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":417572,"visible":true,"origin":"","legend":"","description":"","filename":"Fig02HinetFMD.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/457ff5e596211ab635997cc4.pdf"},{"id":98817703,"identity":"b0f8055e-8d5f-4436-a817-f0544e3e0f81","added_by":"auto","created_at":"2025-12-22 16:39:09","extension":"pdf","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":142261,"visible":true,"origin":"","legend":"","description":"","filename":"Fig03labeldepFMD.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/28c00118f99978d9e9cfeb4f.pdf"},{"id":99307217,"identity":"22b7150a-8696-4751-9963-347872b959a2","added_by":"auto","created_at":"2025-12-31 16:05:49","extension":"pdf","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":37574,"visible":true,"origin":"","legend":"","description":"","filename":"Fig04.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/e26b78cb77ce6f373654c6b9.pdf"},{"id":98817671,"identity":"746400b1-2782-43df-b5f8-a19a3c0c87a3","added_by":"auto","created_at":"2025-12-22 16:39:08","extension":"pdf","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":445606,"visible":true,"origin":"","legend":"","description":"","filename":"Fig05learningcurve.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/e086c24153455a24ed5dda4a.pdf"},{"id":98817676,"identity":"c25de2fb-2861-4558-992e-97c12d9a7118","added_by":"auto","created_at":"2025-12-22 16:39:08","extension":"pdf","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":514309,"visible":true,"origin":"","legend":"","description":"","filename":"Fig06DilatedCov.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/541c8b20641efe106d2c1b4d.pdf"},{"id":99307202,"identity":"f481cf2e-ca4d-4880-af1e-ca01d8039b44","added_by":"auto","created_at":"2025-12-31 16:05:47","extension":"pdf","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":473837,"visible":true,"origin":"","legend":"","description":"","filename":"Fig07DwConv.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/93b8c4c25452a3ae3bfcb1b5.pdf"},{"id":99307301,"identity":"3145985e-462d-4dfb-b090-acceded7ba26","added_by":"auto","created_at":"2025-12-31 16:05:56","extension":"pdf","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":557245,"visible":true,"origin":"","legend":"","description":"","filename":"Fig08InvResNetandConvNeXt.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/0b92f725d8cbc91576989aa3.pdf"},{"id":98817702,"identity":"905f8b92-f2d5-4aa6-9586-7b18969ad5ae","added_by":"auto","created_at":"2025-12-22 16:39:09","extension":"pdf","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":442603,"visible":true,"origin":"","legend":"","description":"","filename":"Fig09KkAperformance.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/530dfd54de4ac406ae082c8d.pdf"},{"id":99307375,"identity":"a673250a-8dac-474f-a23d-ea1a74799e86","added_by":"auto","created_at":"2025-12-31 16:06:05","extension":"pdf","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":398223,"visible":true,"origin":"","legend":"","description":"","filename":"Fig10learningcurve.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/d6c0f7ccb840c79f5a1e6cbb.pdf"},{"id":98817700,"identity":"e8044035-cfcb-4698-be09-8cafb7b311ad","added_by":"auto","created_at":"2025-12-22 16:39:09","extension":"pdf","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":427605,"visible":true,"origin":"","legend":"","description":"","filename":"Fig11PRcurve.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/8c2f7692cf662f723156a252.pdf"},{"id":98817704,"identity":"a9c0d362-1de5-4ed0-af26-3ab4989089a2","added_by":"auto","created_at":"2025-12-22 16:39:09","extension":"pdf","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2328640,"visible":true,"origin":"","legend":"","description":"","filename":"Fig126hypo.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/ae3efeaa4f6ee7dc2aca5dac.pdf"},{"id":98817674,"identity":"b2f32d54-3676-4a2a-a770-2198543ed795","added_by":"auto","created_at":"2025-12-22 16:39:08","extension":"pdf","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":483745,"visible":true,"origin":"","legend":"","description":"","filename":"Fig13validpicks.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/df94e6eb210f85bd1ff860c7.pdf"},{"id":98817696,"identity":"88488343-d15d-4b8d-8927-ab339ee16b2b","added_by":"auto","created_at":"2025-12-22 16:39:09","extension":"xml","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":138967,"visible":true,"origin":"","legend":"","description":"","filename":"EPSPD25003830structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/9922a9b982145f8fb362153a.xml"},{"id":98817680,"identity":"b63795bf-84c5-4e0b-8859-f9adf48f5fb0","added_by":"auto","created_at":"2025-12-22 16:39:08","extension":"html","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":150976,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/53333508bc95c95683bace39.html"},{"id":98817678,"identity":"9e4ce642-153c-4ce0-a3f3-640af2ea6231","added_by":"auto","created_at":"2025-12-22 16:39:08","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":67382,"visible":true,"origin":"","legend":"\u003cp\u003eFMD of events with different hypocenter determination flags: (a) April 11, 2002–March 31, 2016; (b) April 1, 2016–March 21, 2018; and (c) March 22, 2018–March 31, 2023.\u003c/p\u003e","description":"","filename":"fig1.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/2838ac0bc9864d1d1a999cf4.png"},{"id":98817689,"identity":"c830b818-b1b9-4910-a7bd-1d892c640cbe","added_by":"auto","created_at":"2025-12-22 16:39:09","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":33875,"visible":true,"origin":"","legend":"\u003cp\u003eFMD of the Hi-net data sets extracted for model architecture modification: (a) training, (b) validation, and (c) test.\u003c/p\u003e","description":"","filename":"fig2.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/7a8ebb44801dfd49f649b36f.png"},{"id":98817690,"identity":"977630a8-1ba3-43ce-b2ac-d3aed302635a","added_by":"auto","created_at":"2025-12-22 16:39:09","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":69295,"visible":true,"origin":"","legend":"\u003cp\u003eFMD of the Hi-net data sets used to investigate the influence of including k- and A-labeled data in training (Section 4.2): (a) training, (b) validation, and (c) test.\u003c/p\u003e","description":"","filename":"fig3.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/76aa659cd81b393ef8081c24.png"},{"id":99307312,"identity":"b8a329ea-e7d9-428b-85c7-776fcf5de5ce","added_by":"auto","created_at":"2025-12-31 16:05:57","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":137172,"visible":true,"origin":"","legend":"\u003cp\u003eNetwork architectures of the neural phase pickers. (a) Base model (Model 2) developed in this study; (b) Model 7; (c) Models 9 and 10. BN denotes batch normalization.\u003c/p\u003e","description":"","filename":"fig4.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/9a09c38f0595d20216f8fd66.png"},{"id":98817675,"identity":"3bdad8f4-926e-4ad4-a7ef-4efb905cb279","added_by":"auto","created_at":"2025-12-22 16:39:08","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":112172,"visible":true,"origin":"","legend":"\u003cp\u003eLearning curves for the Hi-net training data used to modify neural phase picker models. (a) Models 1–5; (b) Models 2, 4–7; (c) Models 7–12; (d) Models 10, 13–19. The numbers shown in the figure indicate the model numbers. Model 20 was not trained under the same conditions owing to insufficient graphics processing unit (GPU) memory. The results for Models 2, 8, 10, and 19 are shown with thicker lines for clarity.\u003c/p\u003e","description":"","filename":"fig5.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/d7da4df2a5484cd712e48e12.png"},{"id":99307325,"identity":"9502b1d8-73cc-4ff4-bdaa-548650da2b88","added_by":"auto","created_at":"2025-12-31 16:06:00","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":42945,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic diagrams of a dilated convolution and the Atrous Spatial Pyramid Pooling (ASPP) module introduced in Model 7.\u003c/p\u003e","description":"","filename":"fig6.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/62a2f390c534625154ec84cf.png"},{"id":98817687,"identity":"5e96cfa0-e59d-417d-83e8-5e8bd67b2a7c","added_by":"auto","created_at":"2025-12-22 16:39:08","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":25025,"visible":true,"origin":"","legend":"\u003cp\u003eDepthwise separable convolution. (a) Typical 1D convolution. (b) 1D dw-convolution.\u003c/p\u003e","description":"","filename":"fig7.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/2ad28690cab68c18aedb464d.png"},{"id":99307349,"identity":"d4e417c8-19fb-4f07-886c-e4443202667b","added_by":"auto","created_at":"2025-12-31 16:06:04","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":36854,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Inverted residual bottleneck. (b) ConvNeXt block.\u003c/p\u003e","description":"","filename":"fig8.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/8bf905734f76a138c07460b0.png"},{"id":99307726,"identity":"b4205c43-fd2a-43f3-8a94-71eae1d61148","added_by":"auto","created_at":"2025-12-31 16:06:41","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":56910,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance of Model 10 trained on arrival-time data sets with different hypocenter determination flags. PR curves for models trained with different data sets of Hi-net records are shown for (a) P waves and (b) S waves. Thin lines indicate models trained with K labels, thick lines correspond to K + k labels, and dashed lines indicate models trained with K + k + A labels.\u003c/p\u003e","description":"","filename":"fig9.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/2d227b2ed0ea79ec014d1d72.png"},{"id":98817672,"identity":"f0f43994-29a9-4d7f-96ed-83c617cb3044","added_by":"auto","created_at":"2025-12-22 16:39:08","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":38177,"visible":true,"origin":"","legend":"\u003cp\u003eLearning curves obtained during the training of Model 2 (base model), Model 10 (PhaseNeXt-S), Model 19 (PhaseNeXt-M), Model 20 and SegPhase (Katoh et al. 2025) using 25 million records extracted from the JMA unified data set. The dashed lines represent the training loss, and the solid lines represent the validation loss.\u003c/p\u003e","description":"","filename":"fig10.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/270092b7c4e35fe667382c14.png"},{"id":99307215,"identity":"846f9fb7-176f-4502-8311-b34145e02e01","added_by":"auto","created_at":"2025-12-31 16:05:49","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":44218,"visible":true,"origin":"","legend":"\u003cp\u003ePR curves for the base model (Model 2), PhaseNeXt-S (Model 10), PhaseNeXt-M (Model 19), and PhaseNetWC-J developed by Naoi et al. (2024), evaluated on the test data set collected from July 1 to December 31, 2013. (a) P-wave results; (b) S-wave results.\u003c/p\u003e","description":"","filename":"fig11.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/d45e39582889ef5d5cf28abc.png"},{"id":98817699,"identity":"1d22aac2-59f0-41ee-a822-346365c2023a","added_by":"auto","created_at":"2025-12-22 16:39:09","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":175965,"visible":true,"origin":"","legend":"\u003cp\u003eHypocenter distributions on December 31, 2023. (a) JMA unified catalog; (b) Auto-PF catalog; (c) combination of the PF method and PhaseNet-J; (d) combination of the PF method and PhaseNetWC-J; (e) combination of the PF method and PhaseNeXt-S (Model 10); (f) combination of the PF method and PhaseNeXt-M (Model 19). Panels (a) and (b) are reproduced from the graphical abstract of Naoi et al. (2024).\u003c/p\u003e","description":"","filename":"fig12.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/8725cbe0749600990683d77f.png"},{"id":98817673,"identity":"29503c2d-a042-42eb-be96-e109c63a8ce6","added_by":"auto","created_at":"2025-12-22 16:39:08","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":62960,"visible":true,"origin":"","legend":"\u003cp\u003eCounts of picks that were used for event location (effective picks) shown in Figure 12. (a, b) Cumulative numbers of effective picks for P and S waves, respectively. (c, d) Proportions of effective picks for P and S waves, respectively. The proportions were calculated using bins with a width of 0.1.\u003c/p\u003e","description":"","filename":"fig13.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/e6dd43dd313177aae1df1737.png"},{"id":99322307,"identity":"89ffdcf1-6da1-45b7-badf-fa3ae4890124","added_by":"auto","created_at":"2025-12-31 16:43:23","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2121542,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/b30df190-2709-4823-9bb6-0cc452e42534.pdf"},{"id":98817662,"identity":"99273ebd-78d9-4735-a0f8-09796372ca4a","added_by":"auto","created_at":"2025-12-22 16:39:07","extension":"pdf","order_by":17,"title":"","display":"","copyAsset":false,"role":"supplement","size":1778047,"visible":true,"origin":"","legend":"","description":"","filename":"251121Summlement.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/667bd46bde2ea3f51ccaa06b.pdf"},{"id":98817691,"identity":"09959d4a-f360-469c-a230-9f02ab302d69","added_by":"auto","created_at":"2025-12-22 16:39:09","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"supplement","size":124395,"visible":true,"origin":"","legend":"","description":"","filename":"FigGA.png","url":"https://assets-eu.researchsquare.com/files/rs-8174647/v1/fd05d1804efed025803a9f2e.png"}],"financialInterests":"","formattedTitle":"PhaseNeXt: Neural phase picker trained on 20-year records to process the JMA-unified data set","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eIn developing an earthquake catalog from seismic waveform records, the performance of phase picking\u0026mdash;that is, reading arrival times of body waves\u0026mdash;is a crucial factor. In recent years, studies on deep-learning-based phase picking (neural phase picking) have become increasingly active, demonstrating performance superior to that of conventional automatic processing methods (Kubo et al. \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Models trained on large data sets have been released and shared publicly (e.g., Naoi et al. \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Suzuki et al. \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) and are now being employed in seismic activity analyses (Kato \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Although these models exhibit a certain level of generalization and perform reasonably well even on data from regions different from those of the training data set (Hara et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), their performance can be further improved when trained with data obtained from the target region and its specific observation stations. Therefore, for data sets that have a large number of users, it is desirable to develop and share models trained directly on those data (Naoi et al. \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn Japan, following the 1995 Kobe earthquake (Mw 6.9), seismic data that had previously been recorded independently by many institutions were integrated and shared. Since then, the Japan Meteorological Agency (JMA), in collaboration with related organizations, has been responsible for reading arrival times and compiling a unified earthquake catalog (hereinafter referred to as the JMA unified catalog). With the introduction of the Hi-net network (High Sensitivity Seismograph Network in Japan; NIED 2019a) in 2000, earthquake detection capability in Japan has greatly improved, and the seismic network has continued to expand coverage to offshore areas (Aoi et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). At present, continuous waveform data from approximately 2000 stations are routinely recorded, and both the catalog and associated arrival-time data are publicly available on the JMA website (JMA 2025).\u003c/p\u003e \u003cp\u003eNaoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) retrained PhaseNet (Zhu and Beroza \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2019\u003c/span\u003e)\u0026mdash;one of the most widely used neural phase pickers\u0026mdash;using 6.1\u0026nbsp;million arrival-time readings that had been manually reviewed in detail from the JMA unified catalog covering 2014\u0026ndash;2021. The number of training waveforms was roughly ten times greater than that used in the original PhaseNet, which had been trained on California data (Zhu and Beroza \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). The model developed by Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) exhibited significantly better performance for Japan\u0026rsquo;s routine data than the original PhaseNet. They also developed an enlarged model, PhaseNetWC, in which the number of convolutional-layer channels was doubled, resulting in higher performance. The learning curves from these training processes showed no sign of overfitting, suggesting that models with a larger number of parameters and higher expressiveness\u0026mdash;that is, a greater ability to approximate complex functions\u0026mdash;could potentially achieve better performance.\u003c/p\u003e \u003cp\u003eIn this study, we developed a neural phase picker that achieves higher performance for Japan\u0026rsquo;s routine seismic observation data (hereafter referred to as the JMA unified data set). This study addressed the following four tasks: (1) modifying the model architecture to achieve a more parameter-efficient and scalable design; (2) examining preprocessing (preconditioning) methods for input waveforms; (3) demonstrating that incorporating arrival-time readings that had not undergone detailed manual review can improve performance for small-magnitude events; and (4) training three models with different parameter sizes using 20 years of data (2002\u0026ndash;2012 and 2014\u0026ndash;2022), based on the insights from (1) and (3), and evaluating their performance relative to the models developed by Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The trained models are publicly available in the authors\u0026rsquo; GitHub repository.\u003c/p\u003e"},{"header":"2 Data set preparation","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Data extraction from the JMA unified catalog\u003c/h2\u003e \u003cp\u003eIn constructing the JMA unified catalog, all arrival times used for hypocenter determination were processed manually until March 2016. In April 2016, the JMA introduced an automatic processing system based on the Phase combination Forward search (PF) method (Tamaribuchi et al. \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Tamaribuchi \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Since then, only events exceeding regionally defined magnitude thresholds have undergone detailed manual inspection (hereinafter referred to as a full manual review) and have been manually repicked when necessary. The threshold is set to M\u0026thinsp;=\u0026thinsp;1.7 for shallow inland earthquakes and increases with distance from land, reaching up to M\u0026thinsp;=\u0026thinsp;3.5 for oceanic earthquakes. For events with magnitudes below these thresholds, automatically processed or briefly reviewed results are published. Until March 21, 2018, full reviews for subthreshold events continued only in the Tokai region, but after that, this policy was applied nationwide. Detailed explanations of these procedures are provided in Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) and in the JMA (2025) User\u0026rsquo;s Guide.\u003c/p\u003e \u003cp\u003eIn the JMA unified catalog, various labels are assigned to each event and to each arrival-time record (JMA 2025), allowing data to be selectively extracted. Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) selected training data based on the data information flags, which are assigned to each arrival-time reading at individual stations. Each flag consists of a single alphabetic character, with 48 possible patterns in total. Of these, the 24 capital letters denote high-confidence readings that were fully reviewed. Before the introduction of the PF method, all readings were labeled with capital letters, and lowercase flags\u0026mdash;representing lower-confidence readings\u0026mdash;were introduced thereafter. Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) trained PhaseNet and PhaseNetWC using only data associated with capital-letter flags. However, for networks newly deployed after the introduction of the PF method\u0026mdash;such as S-net (the Seafloor Observation Network for Earthquakes and Tsunamis along the Japan Trench; NIED 2019b)\u0026mdash;high-confidence picks were unavailable for small events below the regional magnitude thresholds, possibly resulting in degraded model performance for such events. This issue and our approach to address it are discussed in Sections 2.3 and \u003cspan refid=\"Sec15\" class=\"InternalRef\"\u003e4.2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eData selection based on the data information flags can alternatively be performed more simply using the hypocenter determination flag assigned to each event. This label consists of eight alphabetic characters (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). In this study, we analyzed only events labeled K, which represent the highest quality, and those labeled A or k, which indicate the next-highest quality (JMA 2025). The frequency-magnitude distribution (FMD) of events with these labels is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Before April 2016, no events with k or A labels existed. After the introduction of the PF method in April 2016, the number of small events (M\u0026thinsp;\u0026lt;\u0026thinsp;1.7) labeled K decreased markedly and became nearly absent after March 22, 2018. Events labeled k or A are mostly limited to small magnitudes.\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\u003eHypocenter determination flags\u003c/p\u003e \u003cdiv class=\"Credit\"\u003e\u003cp\u003e(adapted from JMA 2025).\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFlag\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003edescription\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eK\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHigh-precision hypocenters (Manual, closely examined)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLow-precision hypocenters (Manual, closely examined)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ek\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMiddle-precision hypocenters (Manual)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003es\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLow-precision hypocenters (Manual)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMiddle-precision hypocenters (Auto)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ea\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLow-precision hypocenters (Auto)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUndetermined or not accepted or fixed hypocenters\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFar field\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 \u003cp\u003eIn principle, arrival-time readings of events labeled K are accompanied by capital-letter data information flags. However, during a short exceptional period in late November 2016, some K-labeled events were associated with lowercase flags. This occurred because aftershocks following the Mw 7.4 Fukushima-oki earthquake on November 22, 2016, were processed using a procedure different from the routine one. These exceptions account for only 0.04% of all K-labeled data.\u003c/p\u003e \u003cp\u003eThe k and A labels were introduced after the implementation of the PF method. For events below the threshold magnitude for a full manual review, k indicates that a simplified, brief review was conducted, whereas A indicates that the automatic result derived by the PF method was retained without manual inspection. The arrival-time readings of events with k or A labels are generally accompanied by lowercase data information flags. However, a small number of exceptions exist: some k-labeled events that occurred in April\u0026ndash;May 2016 (after the Kumamoto earthquake) and in late November 2016 (after the Fukushima-oki earthquake) have capital-letter flags. These account for only 0.2% of all k-labeled events.\u003c/p\u003e \u003cp\u003eBased on these definitions and their actual usage, it is practical to select data using the K, k, and A labels instead of the data information flags. However, in this study, we followed the original labeling policy: data with K label were used only when accompanied by capital-letter data information flags, and data with the k or A labels only when accompanied by lowercase flags, even when not explicitly mentioned in this paper. Hereafter, we refer to these data as follows: arrival-time readings of events associated with K labels (with capital-letter flags) are termed \u0026ldquo;fully reviewed picks\u0026rdquo;; those associated with k labels (with lowercase flags) are termed \u0026ldquo;briefly reviewed picks\u0026rdquo;; and those associated with A labels are termed \u0026ldquo;automatically read picks.\u0026rdquo; Because the proportion of records with exceptional flags is extremely small, this treatment in training data selection is unlikely to affect the resultant model performance. The cumulative numbers of arrival-time readings and corresponding waveforms extracted under these conditions for events labeled K, k, and A are shown in Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e, and the corresponding hypocenter distributions are shown in Figure \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eAs described in Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), the JMA catalog also includes events other than ordinary earthquakes, such as low-frequency earthquakes, which can be identified through the \u0026ldquo;subsidiary information\u0026rdquo; label assigned to each event. In this study, we extracted and analyzed only events labeled as ordinary earthquakes, whereas Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) used events of all subsidiary information types.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Data selection for model development\u003c/h2\u003e \u003cp\u003eIn Section \u003cspan refid=\"Sec5\" class=\"InternalRef\"\u003e3\u003c/span\u003e, we modify the model architecture of the neural phase picker to achieve higher expressiveness and improve parameter efficiency. To enable repeated training and testing while modifying the model architecture, we prepared a reduced-size data set with a limited number of samples to shorten the computation time required for analysis. In this section, we describe the procedure used to construct this data set.\u003c/p\u003e \u003cp\u003eThe JMA unified data set consists of records from multiple observation networks, including ocean-bottom and broadband networks. As shown by Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), models trained on these data exhibit network-dependent performance. Therefore, when the proportion of data from each network varies within the data set, it becomes difficult to evaluate model performance consistently. To avoid this issue, we used only data from Hi-net, which provides long-term, stable, and densely distributed observations, to examine the model architecture.\u003c/p\u003e \u003cp\u003eThe data used for model development were extracted according to the following three criteria: (1) records from Hi-net stations; (2) events labeled K; and (3) availability of three-component waveform records. To suppress the influence of temporal changes in Japan\u0026rsquo;s routine network and the uneven distribution of large earthquakes and their aftershocks, the data were selected as follows: training data were extracted from 2002, 2005, 2008, 2011, 2014, 2017, 2020, and 2023; validation data from 2003, 2006, 2009, 2012, 2015, 2018, and 2021; and test data from 2004, 2007, 2010, 2013, 2016, 2019, and 2022. We avoided random extraction across the entire period because aftershock sequences, which often contain numerous events occurring within short intervals and confined regions, may produce highly similar waveforms. Such data could cause significant information leakage among the training, validation, and test data sets.\u003c/p\u003e \u003cp\u003eIn addition, to mitigate the imbalance caused by the Gutenberg\u0026ndash;Richter law (Gutenberg and Richter \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e1944\u003c/span\u003e), in which the proportion of small-magnitude events becomes extremely large, we limited the number of samples extracted for each 0.1-magnitude bin to a maximum of 2000 for the training and test data sets and 400 for the validation data set. After applying these conditions, the final data sets contained 123,467 training samples, 26,939 validation samples, and 117,133 test samples. The FMD of these data is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Between magnitudes \u0026minus;\u0026thinsp;0.5 and 5.0, the number of waveforms is approximately uniform across magnitudes, ensuring a well-balanced data set.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003e2.3 Data selection to investigate the influence of using picks not subjected to full review as training data\u003c/b\u003e \u003c/p\u003e \u003cp\u003eIn Section \u003cspan refid=\"Sec15\" class=\"InternalRef\"\u003e4.2\u003c/span\u003e, we evaluate the influence of using the k and A hypocenter determination flags\u0026mdash;introduced after the implementation of the PF method\u0026mdash;for model training. To examine this effect, models trained with data sets compiled after the introduction of the PF method were evaluated using test data from before its implementation. We prepared the training data set from events in January\u0026ndash;December 2019, the validation data set from October\u0026ndash;December 2018, and the test data set from January\u0026ndash;December 2015. From the data obtained in each of these periods, we constructed three data sets labeled K, k, and A, respectively, by extracting records that satisfied the following conditions: (1) records from Hi-net stations and (2) availability of three-component waveform records. The number of samples extracted under each condition is summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, and their magnitude distributions are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. The test data include records from reliably K-labeled events down to M\u0026thinsp;~\u0026thinsp;\u0026minus;\u0026thinsp;1, whereas in the training and validation data sets, records from K-labeled events are largely absent for small magnitudes and are almost entirely replaced by those from k- and A-labeled events.\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\u003eBreakdown of the training, validation, and test data sets used to investigate the influence of including k- and A-labeled data for training (Section \u003cspan refid=\"Sec15\" class=\"InternalRef\"\u003e4.2\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\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 \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHypocenter determination flag\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eValidation\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTest\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eK\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e190,506\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e54,154\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e761,971\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ek\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e117,796\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e26,080\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e655,622\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e188,481\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0\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 \u003c/div\u003e"},{"header":"3 Modification of the architecture of the neural phase picker","content":"\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Features of the PhaseNet architecture\u003c/h2\u003e \u003cp\u003eIn this section, we progressively refine the original PhaseNet architecture (268,443 trainable parameters; Figure S3a) to develop a more parameter-efficient and scalable model that can be extended to larger-scale architectures, thereby achieving greater expressiveness. In this context, expressiveness corresponds to how effectively and how quickly the model can minimize the loss function during training.\u003c/p\u003e \u003cp\u003eA typical approach to enhancing the expressiveness of deep learning models is to increase the number of convolutional layers or the number of filters (channels), as demonstrated by Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), who developed the PhaseNetWC model (1,070,899 trainable parameters). In the field of computer vision, however, various techniques have been developed to construct models with comparable expressiveness while reducing the number of trainable parameters, thereby improving parameter efficiency. Moreover, simply stacking additional layers can often impede training progress due to the vanishing-gradient problem, but techniques that mitigate vanishing gradients even in very deep networks are now well established (He et al. \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). In this study, we incorporate these techniques into the model architecture.\u003c/p\u003e \u003cp\u003eAs pointed out by Zhu and Beroza (\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), the phase-picking problem can be regarded as a one-dimensional (1D) semantic segmentation task. They developed PhaseNet, which adapts U-Net (Ronneberger et al. \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2015\u003c/span\u003e)\u0026mdash;a well-established architecture for semantic segmentation\u0026mdash;to seismic waveform data. PhaseNet has been shown to achieve competitive or superior performance compared with other deep-learning-based phase pickers (M\u0026uuml;nchmeyer et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Garc\u0026iacute;a et al. \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) and has been widely used to construct high-resolution seismic event catalogs (e.g., Kato \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Tan et al. \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). As shown in Figure S3(a), PhaseNet consists of an encoder that compresses the time-series samples through 1D convolutions with a stride of 4 while increasing the number of channels, and a decoder that restores the time resolution using transposed 1D convolutions. The encoder and decoder are connected by skip connections, allowing high-resolution features lost during encoding to be recovered during decoding.\u003c/p\u003e \u003cp\u003ePhaseNet has several distinctive features. The encoder exhibits an exceptionally high compression ratio: whereas most encoder\u0026ndash;decoder architectures for image segmentation compress feature maps to between 1/8 and 1/32 of the input size (e.g., Ronneberger et al. \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Chen et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), PhaseNet reduces them to 1/256 of the original length. The 3001-sample input waveform used by Zhu and Beroza (\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) was compressed to only 11 samples. Such extreme compression helps capture long-range temporal dependencies but likely degrades the temporal resolution required for precise picking, an effect that is mitigated by the inclusion of skip connections. Another characteristic is that each convolutional layer employs filters with a large kernel size of 7, which likely also helps capture long-range dependencies. In conventional architectures, large receptive fields are typically achieved more efficiently by stacking several smaller convolutions (e.g., kernel size 3), making this design choice distinctive. Nevertheless, the recent success of Vision Transformers (ViT) in computer vision has renewed interest in architectures with large kernels (Ding et al. \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Liu et al. \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2022a\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003eb\u003c/span\u003e). In this study, we take these characteristics of PhaseNet into account when refining the model architecture.\u003c/p\u003e \u003cp\u003eNotably, this study focuses on convolutional neural network (CNN) based architectures derived from PhaseNet, although Vision Transformer\u0026ndash;based approaches have recently achieved remarkable success in semantic segmentation tasks. For ViT-based models, we compare our results with SegPhase, a recently proposed ViT-based phase picker that has been shown to outperform PhaseNet (Katoh et al. \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Development of baseline model\u003c/h2\u003e \u003cp\u003eThe architectures of all models developed in this study are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. As a first step, we modified the PhaseNet architecture (Model 1) into a more manageable form. To simplify sample compression by strided convolution and the concatenation of feature maps, we assumed an input length of 4096 samples, which is a power of two. In the decoder, the transposed convolution layers were replaced with a simpler structure consisting of nearest-neighbor upsampling followed by 1D convolution. Each convolution was zero-padded so that, when no stride was applied, the number of samples remained unchanged before and after convolution. For clarity, layers producing outputs with the same number of samples were defined as belonging to the same \u0026ldquo;level,\u0026rdquo; and the number of channels within each level was kept constant, except where skip connections were applied. The model constructed with these modifications, shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(a), is referred to as the base model (Model 2) in this study.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eArchitectures of the neural phase pickers developed in this study.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\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=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel No.\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003edescription\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo. of blocks\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNo. of channels\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNo. of training parameters\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOriginal PhaseNet \u003c/p\u003e \u003cp\u003e*Removed the unnecessary bias term in the first conv. layer.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e268,435\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBase model (simplified PhaseNet with stride\u0026thinsp;=\u0026thinsp;4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 2, 2, 2, 2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e382,915\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIncreased conv. channels of Model 2.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 2, 2, 2, 2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[16, 32, 64, 128, 256]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1,528,323\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIncreased no. of conv. layers of Model 2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[4, 4, 4, 4, 4]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e689,443\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIncreased no. of conv. layers of Model 2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[6, 6, 6, 6, 6]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e995,971\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSimplified PhaseNet (stride\u0026thinsp;=\u0026thinsp;2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 2, 2, 2, 2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e382,915\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSimplified PhaseNet (stride\u0026thinsp;=\u0026thinsp;2)\u0026thinsp;+\u0026thinsp;ASPP module (inspired by DeepLab v3+), based on Model 6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 2, 2, 2, 2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e309,123\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReplaced conv. layers of Model 7 with dw-conv.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 2, 2, 2, 2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e67,872\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReplaced conv. layers of Model 8 with Inverted Residual Bottleneck (expansion ratio\u0026thinsp;=\u0026thinsp;4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 2, 2, 2, 2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e300,771\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReplaced conv. layers of Model 8 with ConvNeXt Blocks (expansion ratio\u0026thinsp;=\u0026thinsp;4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 2, 2, 2, 2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e288,032\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAdded skip connection to the Level 2 layer of Model 10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 2, 2, 2, 2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e288,400\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAdded skip connection to the Level 1 layer of Model 10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 2, 2, 2, 2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e288,152\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIncreased no. of conv. layers of Model 10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 2, 4, 6, 4]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e704,544\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIncreased no. of conv. layers of Model 10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 3, 3, 9, 3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e666,208\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIncreased no. of conv. layers of Model 10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 3, 3, 27, 3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[8, 16, 32, 64, 128]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1,269,856\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIncreased conv. channels of Model 13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 2, 4, 6, 4]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[12, 24, 48, 96, 192]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1,574,268\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e17\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIncreased conv. channels of Model 13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 2, 4, 6, 4]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[16, 32, 64, 128, 256]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2,788,984\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIncreased conv. channels of Model 14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 3, 3, 9, 3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[12, 24, 48, 96, 192]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1,487,580\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIncreased conv. channels of Model 15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 3, 3, 27, 3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[12, 24, 48, 96, 192]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2,835,420\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIncreased conv. channels of Model 19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e[2, 3, 3, 27, 3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e[16, 32, 64, 128, 256]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e5,021,432\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 \u003cp\u003eIn the original PhaseNet (Model 1), the number of channels doubles in the first convolution of each level (Figure S3a). In the base model (Model 2), however, channel expansion is performed simultaneously with sample compression by stride, resulting in an approximately 1.5-fold increase in trainable parameters (382,915 in total). Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(a) shows the training losses over epochs for Model 1 and Model 2, trained using the data set prepared in Section \u003cspan refid=\"Sec4\" class=\"InternalRef\"\u003e2.2\u003c/span\u003e, with waveforms segmented into 4096-sample windows. To prevent the model from memorizing the arrival-time position, each waveform was extracted so that the P-wave arrival could appear anywhere within the entire window, determined using a uniform random offset. The inclusion of the S wave within the window was not strictly necessary. The ground-truth labels were assigned in the same manner as in Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), using the JMA arrival-time readings as peak positions with a height of 1.0 and representing them by Gaussian functions with a standard deviation of 0.1 s. Each window contained one pair of P and S phases unless the S wave extended beyond the window. This window-positioning and labeling procedure was also applied in subsequent analyses unless otherwise noted. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(a), the increased number of trainable parameters in Model 2 resulted in slightly lower loss values than those of Model 1. Because the purpose of this section is to examine the expressiveness of the model architecture, Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e presents only the training loss. Although not shown here, the validation loss began to diverge from the training loss at around 80 epochs for Model 1 and at around 60 epochs for Model 2, suggesting the onset of overfitting.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAs straightforward extensions to enhance the expressiveness of the base model, we constructed Model 3 by doubling the number of channels in the convolutional layers of Model 2 (1,528,323 trainable parameters), Model 4 by increasing the number of convolutional layers in each encoder level to four (689,443 parameters), and Model 5 by increasing the number to six (995,971 parameters). All other hyperparameters were identical to those of Model 2. The learning curves of these models, shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(a), indicate that Model 3\u0026mdash;the one with the largest number of trainable parameters\u0026mdash;showed the greatest loss reduction; however, compared with the models described later, the reduction in loss relative to the increase in parameter count was less efficient. Improving the efficiency of loss reduction is the focus of the following sections.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Adjustment of the sample-size compression ratio and modification for capturing long-range features\u003c/h2\u003e \u003cp\u003eAs mentioned earlier, both the original PhaseNet (Model 1) and the base model (Model 2) compress seismic waveforms to 1/256 of the input length. For an input of 3001 samples, the feature maps are reduced to 11 samples, and for 4096 samples, to 16 samples. Because this high compression ratio likely sacrifices the resolution required for accurate picking, we reduced the ratio by changing the stride width of each layer from 4 to 2, resulting in a final compression ratio of 1/16.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(b) shows the learning curve obtained for Model 6, in which the stride width was changed to 2 (Figure S3b). Compared with the base model (Model 2), the loss decreased more slowly, suggesting that the model\u0026rsquo;s expressiveness was reduced owing to the limited ability to capture long-range features. This ability to capture long-range dependencies is considered one of the key factors underlying the strong performance of ViTs in segmentation tasks (Khan et al. \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) and in phase-picking problems (Katoh et al. \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Several studies have attempted to introduce the capability to learn long-range characteristics into convolution-based models for segmentation problems. One such approach is dilated convolution (Yu and Koltun \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Yu et al. \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ea), in which convolutional operations are applied at fixed intervals (rates) to expand the receptive field. Before the introduction of ViTs, the DeepLab v3\u0026thinsp;+\u0026thinsp;model (Chen et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) gained popularity for segmentation tasks by introducing the Atrous Spatial Pyramid Pooling (ASPP) module (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eb), which combines the outputs of multiple dilated convolutions performed with different rates to capture features at various spatial scales.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn this study, we incorporated the ASPP module and the decoder of DeepLab v3\u0026thinsp;+\u0026thinsp;to capture long-range information while maintaining stride-2 convolutions. The resulting model architecture (Model 7; 309,123 parameters) is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(b), and its learning curve is superimposed in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(b). Despite having approximately 20% fewer parameters than Model 2, the ASPP-equipped model achieved lower loss values than Model 2 and also than Models 4 and 5, which have a greater number of layers (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Parameter reduction using depthwise separable convolution\u003c/h2\u003e \u003cp\u003eNext, to reduce the number of trainable parameters, we introduced depthwise separable convolution (Sifre \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Chollet \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; hereinafter referred to as dw-convolution). In a standard convolution, filtering is performed with kernels that have trainable parameters across both the spatial and channel dimensions (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003ea). In contrast, dw-convolution first applies a depthwise convolution, in which independent convolutions are performed for each channel, followed by a pointwise convolution using 1\u0026times;1 kernels to combine information across channels (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003eb). This structure enables a substantial reduction in the number of parameters while maintaining the same input\u0026ndash;output dimensions.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eWe constructed Model 8 by replacing the convolutional layers of Model 7 with dw-convolution (Figure S3c). Only the convolution used in the global average pooling (GAP) within the ASPP module, the final convolution in the ASPP module, and the last convolutional layer of the network remained standard convolutions. As a result, the total number of parameters decreased to 67,872, but the training loss increased significantly, indicating reduced expressiveness (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec). This reduction in expressiveness is addressed in the next section by increasing the number of layers based on the inverted residual bottleneck (Sandler et al. \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) and the ConvNeXt block (Liu et al. \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2022b\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Increasing model scalability by introducing inverted residual bottlenecks and ConvNeXt blocks with residual connections\u003c/h2\u003e \u003cp\u003eAs described in Section \u003cspan refid=\"Sec6\" class=\"InternalRef\"\u003e3.1\u003c/span\u003e, simply increasing the number of convolutional layers does not always lead to better performance and can often cause degradation. If additional layers could perfectly behave as an identity mapping, such degradation would not occur; therefore, an important limitation of CNNs is that their layers cannot easily learn identity mappings. To address this issue, He et al. (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2016\u003c/span\u003e) introduced residual connections and developed the ResNet model, which enables stable training of much deeper networks. In this study, we also constructed scalable models that can stack more layers by incorporating residual connections.\u003c/p\u003e \u003cp\u003eHere, we enhance expressiveness by stacking inverted residual bottlenecks (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003ea) and ConvNeXt blocks (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003eb) with residual connections. The inverted residual bottleneck first expands the number of channels in the feature maps using 1\u0026times;1 (pointwise) convolutions, applies a depthwise convolution with multiple channels, and then performs another 1\u0026times;1 convolution to produce output feature maps with the desired number of channels. By placing pointwise convolutions before and after the depthwise convolution, this block achieves high expressiveness while keeping the number of parameters moderate, because the channel dimensions outside the block remain compact. We constructed Model 9 (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ec) using inverted residual bottlenecks that have the same input and output feature-map sizes as the convolutional layers in the encoder parts of Models 7 and 8. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e(a), the channel expansion ratio in the intermediate layer was set to 4, and the kernel size to 7. Model 9 achieved loss values comparable to those of Model 7 (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec), with slightly fewer trainable parameters.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eNext, we constructed Model 10, in which each inverted residual bottleneck was replaced with a ConvNeXt block. The ConvNeXt block reverses the order of the first and second layers in the inverted residual bottleneck and modifies the activation functions (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003eb). These architectural refinements incorporate design insights gained from ViTs, allowing convolution-based networks to achieve performance comparable to ViTs (Liu et al. \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2022b\u003c/span\u003e). The total number of parameters of Model 10 was 4.2% smaller than that of Model 9 (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e), while achieving a lower training loss (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec). In the following sections, we use Model 10 as the base for further improvement. We refer to this type of architecture as PhaseNeXt, and to Model 10 specifically as PhaseNeXt-S.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.6 Skip connection\u003c/h2\u003e \u003cp\u003eIn all models developed so far, only the Level-3 skip connection was passed to the decoder, following the structure of DeepLab v3+. To examine the effect of skip connections from shallower layers, we constructed Model 11, which added a Level-2 skip connection, and Model 12, which added a Level-1 skip connection, to Model 10. Training results showed no noticeable difference in the learning curves (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec) compared with Model 10. Therefore, in this study, skip connections from Levels 1 and 2 were not included. It should also be noted that, whereas the original PhaseNet used stride-4 convolutions, the present models employ stride-2 convolutions. Consequently, the two-step upsampling from Level 3 in Models 6\u0026ndash;12 corresponds to only one step of resolution recovery in the original PhaseNet. This may be the reason why the additional skip connections did not make a significant difference.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.7 Scaling model size by stacking blocks and widening channels\u003c/h2\u003e \u003cp\u003eSince the proposal of ResNet (He et al. \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2016\u003c/span\u003e), it has become common to enhance model expressiveness by stacking multiple standardized blocks, such as inverted residual bottlenecks and ConvNeXt blocks. In such architectures, expressiveness can be easily scaled by adjusting the number of blocks and channels. Following common design practices observed in modern convolutional networks (He et al. \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Sandler et al. \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), fewer blocks are typically used in shallower levels\u0026mdash;where feature maps are larger\u0026mdash;and more blocks are stacked in deeper levels to balance representational power and computational efficiency. Furthermore, as reported by Tan and Le (\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), simultaneously scaling both the depth (number of layers) and the width (number of channels) provides a more efficient way to enhance expressiveness than increasing either the depth or the width alone.\u003c/p\u003e \u003cp\u003eBased on Model 10, we developed and trained a series of larger models (Models 13\u0026ndash;19; Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) by increasing the number of layers (blocks), the number of channels, or both. The corresponding learning curves are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e(d). As the model size increased, the loss decreased, and the minimum loss values were achieved by Models 17 and 19, which contained the largest number of parameters. As demonstrated by Models 3\u0026ndash;5, such low loss values are difficult to achieve simply by increasing the number of channels or layers in the base model, highlighting the importance of the architectural refinements.\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Data selection and preprocessing schemes","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Preprocessing of signal amplitudes\u003c/h2\u003e \u003cp\u003eWhen inputting waveforms into neural phase pickers, normalization or standardization is often applied to each channel individually (Zhu and Beroza \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Such preprocessing helps stabilize and accelerate model training by aligning the scales of input values. However, in seismic waveforms, the amplitude ratio among channels carries important information. For example, the initial motion of P wave is aligned with the ray path, and this linearity may provide a key indicator for identifying the P phase. Therefore, applying normalization that preserves the inter-channel amplitude ratio may improve the performance of phase-picking models.\u003c/p\u003e \u003cp\u003eIn addition, Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) reported that the picking performance of PhaseNet and PhaseNetWC decreased as event magnitude increased. Although they attributed this magnitude dependence to the shortage of large-magnitude events in the training data\u0026mdash;a result of the Gutenberg\u0026ndash;Richter law\u0026mdash;we consider another possible factor: amplitude normalization applied during preprocessing may reduce P-wave picking performance. Specifically, when normalized waveforms from large-amplitude events are fed into the model, the relative change from noise to the P-wave onset becomes smaller within the overall amplitude range, potentially leading to degraded P-wave picking performance for high-magnitude events. To mitigate this issue, normalization based on the noise level with amplitude saturation at a certain threshold or a soft-clipping technique (Zhu et al. \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) could be applied; however, these approaches would inevitably lead to some loss of waveform information. Instead, in this study, we examined whether this degradation in performance could be mitigated by simultaneously inputting six channels\u0026mdash;three fully normalized components and three additional components in which the P-wave portion was amplified using noise-level-based normalization with amplitude clipping.\u003c/p\u003e \u003cp\u003eBased on these ideas, we trained the same model under four input configurations and compared their performance: (1) three-component waveforms standardized independently for each channel (z-score normalization); (2) three-component waveforms standardized while preserving inter-channel amplitude ratios (each channel was mean-corrected, and the three components were jointly normalized by their overall standard deviation); (3) six-component waveforms composed of the three individually standardized components and an additional three P-wave-amplified components; and (4) six-component waveforms composed of the three components standardized while preserving inter-channel amplitude ratios and another three P-wave-amplified components that also preserve inter-channel amplitude ratios. The approaches used by Zhu and Beroza (\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) and Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) correspond to case (1). The amplification of waveforms was implemented by standardizing the first 150 samples of each waveform segment using their own mean and standard deviation. As described in Section \u003cspan refid=\"Sec7\" class=\"InternalRef\"\u003e3.2\u003c/span\u003e, the P-wave arrival was generally positioned at a random location within the whole of the cutting window. However, in this analysis, we prepared the data so that the P-wave arrival did not occur within the first 150 samples.\u003c/p\u003e \u003cp\u003eThis analysis was conducted using the data set prepared in Section \u003cspan refid=\"Sec4\" class=\"InternalRef\"\u003e2.2\u003c/span\u003e, with the Model 10 architecture trained under the four input configurations described above. In cases (3) and (4), the number of input channels was expanded from three to six in the model illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e(c). The models that achieved the minimum validation loss in each training run were used for performance evaluation on the test data. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the number of samples in the data set exhibited no magnitude dependence within the range of approximately \u0026minus;\u0026thinsp;0.5\u0026thinsp;\u0026lt;\u0026thinsp;M\u0026thinsp;\u0026lt;\u0026thinsp;5, allowing performance evaluation without bias arising from the training-data imbalance noted by Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The evaluation followed the approach of Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), in which true positives, false positives, and false negatives were defined according to Sun et al. (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) (Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e), and performance was assessed using precision\u0026ndash;recall (PR) curves.\u003c/p\u003e \u003cp\u003eFigure S4 shows the PR curves for the models trained using the four input configurations. Except for the P-wave results at M\u0026thinsp;\u0026lt;\u0026thinsp;0 (whose lower performance likely reflects the lower signal-to-noise ratio of the waveforms, as noted by Naoi et al. \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), a magnitude-dependent trend similar to that observed by Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) remained evident for both P and S phases, even though the training data set was designed to minimize magnitude dependence. This tendency did not improve even when the models were trained using amplified waveforms (cases 3 and 4) or when inter-channel amplitude ratios were preserved (cases 2 and 4). These results indicate that the degradation of picking performance with increasing magnitude is not caused by deficiencies in the preprocessing or composition of the training data, but rather reflects an intrinsic property of the waveforms themselves. Because the differences in performance were minor, subsequent analyses were conducted using configuration (1)\u0026mdash;the simplest approach, consistent with that adopted by Zhu and Beroza (\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) and Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Incorporating briefly reviewed and automatically picked data into training\u003c/h2\u003e \u003cp\u003eAs described in Section \u003cspan refid=\"Sec3\" class=\"InternalRef\"\u003e2.1\u003c/span\u003e, picks associated with K-labeled events (i.e., events subjected to full manual review) were lacking for small-magnitude events after the introduction of the PF method, likely resulting in reduced performance for small events recorded by the networks developed thereafter. To examine whether this performance degradation could be mitigated by incorporating briefly reviewed or automatically processed picks into the training data set, we trained Model 10 using three data sets derived from the Hi-net data prepared in Section 2.3: (1) data associated with K-labeled events, (2) data associated with either K- or k-labeled events, and (3) data associated with K-, k-, or A-labeled events.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e shows the PR curves obtained using Model 10 at the epoch that achieved the minimum validation loss in the training with each of the three data sets. The test data consisted of events from 2015, all labeled as K. The model trained with data set (1) exhibited lower detection performance for small events, particularly those with M\u0026thinsp;\u0026lt;\u0026thinsp;0, reflecting the shortage of such small-magnitude events in its training data. In contrast, training with data sets (2) and (3) substantially improved detection performance for these small events. These results indicate that combining briefly reviewed and automatically read picks in the training process can compensate for the lack of K-labeled data in recently deployed observation networks. Although data set (3) contained more training samples than data set (2), its performance deteriorated, particularly for very small events. In the following analyses, we use the data set combining K and k labels for model training.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"5 Training of modified models using a large JMA unified data set","content":"\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Learning curves\u003c/h2\u003e \u003cp\u003eBased on the findings obtained in Sections \u003cspan refid=\"Sec5\" class=\"InternalRef\"\u003e3\u003c/span\u003e and \u003cspan refid=\"Sec13\" class=\"InternalRef\"\u003e4\u003c/span\u003e, we trained neural phase pickers using arrival-time records listed in the JMA unified catalog for 2002\u0026ndash;2012 and 2014\u0026ndash;2023, along with the corresponding waveforms extracted from continuous records. For validation and testing, we used the 2013 data set, in which all picks were fully reviewed: data from January to June were used for validation, and those from July to December for testing. As described in the previous section, events with hypocenter determination flags of K and k were used for training. The total number of waveforms used for training was 24,469,146 (22,610,301 with K flags and 1,858,845 with k flags), approximately four times larger than that used by Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), with 820,328 waveforms for validation and 1,229,716 for testing (all with K flags). The 2013 data were also used as test data in Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) and were not included in their training. Therefore, using the same test data set, we compared the performance of the models trained in this study with those trained by Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). We refer to the PhaseNet retrained by Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) using the eight-year JMA data set as PhaseNet-J, and to the PhaseNetWC trained by them as PhaseNetWC-J.\u003c/p\u003e \u003cp\u003eUsing this data set, we trained three models for 60 epochs under the same computational environment and training parameters: the base model (Model 2), PhaseNeXt-S (Model 10), and PhaseNeXt-M (Model 19). PhaseNeXt-S has about 75% of the trainable parameters of the base model, whereas PhaseNeXt-M has approximately 7.4 times more. Training was conducted using two NVIDIA A40 GPUs. With a batch size of 1024, the training times were 119 h for the base model, 126 h for PhaseNeXt-S, and 398 h for PhaseNeXt-M.\u003c/p\u003e \u003cp\u003eFor comparison, we also trained SegPhase, a ViT-based model proposed by Katoh et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), and Model 20, in which the number of channels in PhaseNeXt-M was further increased. For SegPhase, we adopted the convolutional stride settings specified by Katoh et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; their Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e), with values of (st₁, st₂, st₃) = (2, 2, 2), and the total number of trainable parameters in this configuration was 333,107. Because of their substantial computational demands, these two models could not be trained under the same environment or batch-size settings as the previous three models. SegPhase was trained for 20 epochs on the same system with a reduced batch size of 512, which took 300 h; extrapolation suggests that approximately 648 h would have been required to train it for 60 epochs. Model 20 was trained on the JMA supercomputer equipped with eight NVIDIA A100 GPUs, using a batch size of 4096 for 20 epochs.\u003c/p\u003e \u003cp\u003eThe resulting learning curves are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e. As expected from the training-loss differences shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, both the training and validation losses decreased successively from the base model to PhaseNeXt-S and PhaseNeXt-M. Notably, in all cases, the validation loss was smaller than the training loss\u0026mdash;opposite to the trend typically observed in deep-learning training. This suggests that the validation data were easier for the model to pick than the training data. This tendency is likely because the validation data consisted exclusively of K-labeled events from 2013.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eEven for PhaseNeXt-M, no signs of overfitting were observed, similar to the findings reported by Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) for PhaseNet and PhaseNetWC, suggesting that further scaling of the model size may lead to improved performance. Indeed, in training of the larger model (Model 20), the minimum validation loss was slightly smaller than that of PhaseNeXt-M; however, such large models require substantially longer training and prediction times. Although further enlarging the model may yield additional performance gains, practical applications would require model pruning or other post-training compression techniques to reduce computational costs.\u003c/p\u003e \u003cp\u003eRegarding SegPhase, although it has a comparable number of trainable parameters to the base model, its training cost was even higher than that of PhaseNeXt-M due to the inclusion of multi-head attention, whereas its performance only slightly exceeded that of PhaseNeXt-S and remained lower than that of PhaseNeXt-M. While ViT-based architectures may demonstrate superior performance to CNN-based models when trained on extremely large data sets with sufficient computational resources\u0026mdash;as achieved in segmentation tasks and large language models\u0026mdash;the advantage of ViT-based approaches was not evident under the data volume and computational environment considered in this study.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e5.2 PR curves for the 2013 test data set\u003c/h2\u003e \u003cp\u003eFor the base model, PhaseNeXt-S, and PhaseNeXt-M, we adopted the results corresponding to the minimum validation losses (epoch 20 for the base model and epoch 53 for both PhaseNeXt-S and PhaseNeXt-M) in the subsequent analysis. The resultant models, along with the PhaseNetWC-J model developed by Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), were evaluated using the test data from July\u0026ndash;December 2013. The PR curves for the test data set are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e. The performances of the base model and PhaseNetWC-J were nearly identical, whereas clear improvements\u0026mdash;indicated by PR curves shifting toward the upper-right corner\u0026mdash;were observed for PhaseNeXt-S and PhaseNeXt-M. Models that achieved lower loss values in the training curves shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e tended to exhibit better performance. It should be noted that the 2013 test data consisted of records acquired before the deployment of DONET2 (Dense Ocean floor Network for Earthquakes and Tsunamis; NIED 2019c) and S-net, for which the influence of incorporating k-labeled events is expected to be more pronounced.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Comparison of application results for the JMA unified data set\u003c/h2\u003e \u003cp\u003eNaoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) compared the hypocenter determination results obtained using the JMA automatic processing routine with those from a workflow in which the picking method was replaced by their retrained PhaseNet-J, based on data recorded on March 31, 2023, after adjusting the P- and S-phase thresholds to yield approximately the same number of picks. In this study, we applied the same approach to PhaseNetWC-J, PhaseNeXt-S, and PhaseNeXt-M to examine whether the present models outperform PhaseNetWC-J. The numbers of picks and determined hypocenters are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, and the resulting hypocenter distributions are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e. The \u0026ldquo;Auto-PF\u0026rdquo; result shown in the table and figure refers to the catalog developed using the JMA-adopted automatic processing algorithm. Although the PhaseNet-J results shown for comparison differ slightly from those reported by Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), this discrepancy arises because Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) applied an additional process to read amplitudes and repick arrival times using the same method employed in the development of the JMA unified catalog (this repicking process slightly adjusted the PhaseNet-J arrival times, and when repicking failed, the original PhaseNet-J picks were retained). Because this study did not apply such post-processing, the comparison is more direct; however, this difference has only a minor influence on the overall discussion. PhaseNeXt-S determined nearly the same number of events as PhaseNetWC-J, whereas PhaseNeXt-M achieved the largest number of determined hypocenters\u0026mdash;approximately 3.5 times as many as those listed in the JMA unified catalog.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eNumbers of picks and cataloged events obtained from continuous data recorded on December 31, 2023 (updated from Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e in Naoi et al. \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\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=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u003csup\u003e\u003cem\u003eth\u003c/em\u003e\u003c/sup\u003e for P-wave\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e\u003csup\u003e\u003cem\u003eth\u003c/em\u003e\u003c/sup\u003e for S-wave\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNo. of P-picks\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNo. of S-picks\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo. of hypocenters\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMatched\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnified catalog\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e636\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAuto-PF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e427,436\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e735,105\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e911\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e521(2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePhaseNet-PF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.250\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.106\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e436,107\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e735,539\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1747\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e561(3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePhaseNetWC-PF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.310\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.090\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e432,814\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e736,685\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1909\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e550(4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePhaseNeXt-S\u003c/p\u003e \u003cp\u003e(Model 10)-PF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.234\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.031\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e427,283\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e729,928\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1878\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e552(4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePhaseNeXt-M (Model 19)-PF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.200\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e430,500\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e758,514\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2135\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e564(4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"7\"\u003e* \u0026lsquo;Matched' indicates that the corresponding events were found in the JMA unified catalog (Naoi et al. \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The numbers shown in brackets represent duplicate counts (i.e., cases where multiple detected events are associated with a single event in the unified catalog).\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn the above analysis, the PhaseNeXt models required significantly lower peak thresholds compared with PhaseNetWC-J, especially for the S wave, to obtain approximately the same number of picks. To evaluate how these picks with low peak values contributed to hypocenter determination, Figs.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(a) and 13(b) show the cumulative number of effective picks\u0026mdash;defined as those actually used for hypocenter determination\u0026mdash;whereas Figs.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(c) and 13(d) present the proportion of effective picks, closely corresponding to the \u0026ldquo;utilization rate\u0026rdquo; defined by Katoh et al. (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), for each peak-value bin with an interval of 0.1. From Figs.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(a) and 13(b), it is evident that the number of effective picks obtained with the PhaseNeXt models increased relative to those obtained with PhaseNet-J and PhaseNetWC-J. For PhaseNeXt-M, the advantage was found for peak values below 0.8 for the P phase and between 0.2 and 0.8 for the S phase. For PhaseNeXt-S, a similar number of effective picks to those obtained with PhaseNeXt-M was found in the same range for the S wave.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(b) indicates that S-phase picks with very low peak values below 0.1 significantly contribute to the large number of effective picks obtained with the PhaseNeXt models. When the same threshold as that applied to PhaseNetWC-J was used for PhaseNeXt-M, the number of available picks decreased to approximately 42% (269,317 for P waves and 231,342 for S waves, totaling 500,659 picks). Nevertheless, the number of determined hypocenters remained identical to that obtained with PhaseNetWC-J (1909 events). For PhaseNeXt-S, using the same threshold resulted in 92% of the hypocenters (1751 events) being determined from 54% of the picks (308,545 for P waves and 321,507 for S waves, totaling 630,052 picks). These results clearly demonstrate the higher precision of the PhaseNeXt models compared with PhaseNetWC-J.\u003c/p\u003e \u003cp\u003eFigures \u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e(c) and 13(d), which show the percentage of effective picks, directly demonstrate the high precision of the PhaseNeXt models. This advantage is particularly evident for the S phase, where PhaseNeXt-S and PhaseNeXt-M clearly outperform PhaseNet-J and PhaseNetWC-J. Although this characteristic is favorable for earthquake catalog compilation, in networks with many stations and highly accurate phase association, even a large number of false picks has only a minor effect on the final number of determined hypocenters. This likely explains the relatively small difference in the number of determined events among PhaseNetWC-J, PhaseNeXt-S, and PhaseNeXt-M shown in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. The advantage of the PhaseNeXt models is expected to become more pronounced in cases where the number of stations is limited and false detections have a greater impact.\u003c/p\u003e \u003cp\u003eIn any case, PhaseNeXt-M exhibited the best performance among all models compared. However, owing to its larger architecture, additional computational cost is required not only for training but also for prediction. In our test applying the model to continuous waveform data, PhaseNeXt-M required approximately 4.3 times more computation time than PhaseNetWC-J. This additional cost in picking may become a constraint in near\u0026ndash;real-time catalog development or large-scale applications, although the reduced number of false detections produced by PhaseNeXt-M helps to lower the computational cost of the subsequent phase association process. For prediction, PhaseNeXt-S required approximately 1.3 times more computation time than PhaseNetWC-J, even though PhaseNeXt-S has fewer parameters than PhaseNetWC-J. Its computational load appears to have increased due to architectural differences. As is often the case with large language models, practical applications may require selecting an appropriate model by considering the trade-off between computational cost and accuracy.\u003c/p\u003e \u003c/div\u003e"},{"header":"6 Discussion and conclusions","content":"\u003cp\u003eIn this study, we extended the work of Naoi et al. (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) on training deep-learning-based phase pickers using the JMA unified data set by incorporating seismic records spanning 2002\u0026ndash;2023. From more than 25\u0026nbsp;million waveform records, we developed and trained a new neural phase picker, PhaseNeXt. PhaseNeXt builds on PhaseNet but integrates recent advances in computer vision\u0026mdash;specifically, the ConvNeXt block and the Atrous Spatial Pyramid Pooling (ASPP) module\u0026mdash;to achieve higher expressiveness and greater parameter efficiency. During training, we demonstrated that including briefly reviewed picks, in addition to those subjected to detailed manual inspections, further improved performance. The final model, PhaseNeXt-M, outperformed the existing PhaseNet-J and PhaseNetWC-J models and, when combined with the PF method, detected approximately 3.5 times more hypocenters than those listed in the JMA unified earthquake catalog, while using a comparable number of picks to the JMA automatic processing.\u003c/p\u003e \u003cp\u003eAlthough PhaseNeXt-M achieved high performance, several challenges remain. To further improve performance and facilitate practical implementation, the following directions can be considered: (1) enhancing data augmentation: training with techniques such as noise addition (Naoi et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) and waveform superposition (Kim et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) may improve detection performance under low signal-to-noise conditions or during periods of frequent seismicity; (2) mitigating label imbalance: phase picking is inherently a highly imbalanced classification problem, as most labels correspond to noise. As pointed out by Katoh et al. (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), introducing weighted learning could improve both training efficiency and model performance; (3) model compression: applying techniques such as pruning and knowledge distillation may enable the development of lightweight models capable of real-time processing without significant loss of accuracy; and (4) improving performance for large events. Advancing these improvements and reconstructing earthquake catalogs from the extensive continuous waveform archives accumulated to date could further enhance seismic monitoring systems, and the integration of PhaseNeXt into such systems could lead to more accurate and comprehensive earthquake catalogs in the future.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eJMA Japan Meteorological Agency\u003c/p\u003e\n\u003cp\u003ePF Phase combination Forward search\u003c/p\u003e\n\u003cp\u003eCNN Convolutional neural network\u003c/p\u003e\n\u003cp\u003eViT Vision Transformer\u003c/p\u003e\n\u003cp\u003eGPU Graphics Processing Unit\u003c/p\u003e\n\u003cp\u003eASPP Atrous Spatial Pyramid Pooling\u003c/p\u003e\n\u003cp\u003eBN Batch Normalization\u003c/p\u003e\n\u003cp\u003edw-convolution Depthwise separable convolution\u003c/p\u003e\n\u003cp\u003eGAP Global average pooling\u003c/p\u003e\n\u003cp\u003ePR Precision\u0026ndash;recall\u003c/p\u003e\n\u003cp\u003eFMD Frequency magnitude distribution\u003c/p\u003e"},{"header":"Declarations","content":"\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\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe base model, PhaseNeXt-S, and PhaseNeXt-M models trained using the 20-year JMA unified data set are available at https://github.com/mktnaoi/JMAuniPicker. Seismic waveform data related to the routine seismic observations in Japan (including a network other than Hi-net) are available at http://www.hinet.bosai.go.jp/?LANG=en. The JMA unified catalog is available at https://www.data.jma.go.jp/svd/eqev/data/bulletin/index.html.\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\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by JSPS KAKENHI (Grant Nos. JP21H01191, JP20K14565, JP24H01027, and JP25K01083) and by the Ministry of Education, Culture, Sports, Science, and Technology (MEXT) of Japan under its Earthquake and Volcano Hazards Observation and Research Program.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMN was the primary contributor to analysis and manuscript writing. KS and KT provided detailed information on handling the JMA unified data set and contributed to catalog development and figure preparation. All authors contributed to the writing and have read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study used data deriving from the National Research Institute for Earth Science and Disaster Resilience (2019a\u0026ndash;d), Hokkaido University, Hirosaki University, Tohoku University, the University of Tokyo, Nagoya University, Kyoto University, Kochi University, Kyushu University, Kagoshima University, National Institute of Advanced Industrial Science and Technology (AIST), Geospatial Information Authority of Japan, Japan Agency for Marine-Earth Science and Technology (JAMSTEC), Aomori Prefecture, Tokyo Metropolitan Government, Shizuoka Prefecture, Yokohama City (Kanagawa Prefecture), Hot Springs Research Institute of Kanagawa Prefecture, Association for the Development of Earthquake Prediction, Group for Urgent Joint Seismic Observation of the 2016 Kumamoto Earthquake, and Japan Meteorological Agency. Regarding the analysis, we employed a large continuous seismic data analysis system (Nakagawa et al. 2016) at the Earthquake Research Institute, University of Tokyo (ERI JURP 2024-F3-12). We used SeisBench (Woollam et al. 2022) based on the PyTorch library for deep-learning analysis.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eAoi S, Asano Y, Kunugi T et al (2020) MOWLAS: NIED observation network for earthquake, tsunami and volcano. Earth Planets Space 72:126. https://doi.org/10.1186/s40623-020-01250-x\u003c/li\u003e\n \u003cli\u003eChen LC, Zhu Y, Papandreou G et al (2018) Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European Conference on Computer Vision. Springer, pp 833\u0026ndash;851. https://doi.org/10.1007/978-3-030-01234-2_49\u003c/li\u003e\n \u003cli\u003eChollet F (2017) Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, pp 1800-1807. https://doi.org/10.1109/CVPR.2017.195\u003c/li\u003e\n \u003cli\u003eDing X, Zhang X, Han J, Ding G (2022) Scaling up your kernels to 31\u0026times;31: Revisiting large kernel design in CNNs. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, pp 11963\u0026ndash;11975. https://doi.org/10.1109/CVPR52688.2022.01166\u003c/li\u003e\n \u003cli\u003eGarc\u0026iacute;a JE, Fern\u0026aacute;ndez-Prieto LM, Villase\u0026ntilde;or A et al (2022) Performance of deep learning pickers in routine network processing applications. Seismol Res Lett 93:2529\u0026ndash;2542. https://doi.org/10.1785/0220210323\u003c/li\u003e\n \u003cli\u003eGutenberg B, Richter CF (1944) Frequency of earthquakes in California. Bull Seism Soc Am 34:185\u0026ndash;188. https://doi.org/10.1785/BSSA0340040185\u003c/li\u003e\n \u003cli\u003eHara S, Fukahata Y, Iio Y (2019) P-wave first-motion polarity determination of waveform data in western Japan using deep learning. Earth Planets Space 71:1\u0026ndash;11. https://doi.org/10.1186/s40623-019-1111-x\u003c/li\u003e\n \u003cli\u003eHe K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE. https://doi.org/10.1109/CVPR.2016.90\u003c/li\u003e\n \u003cli\u003eJapan Meteorological Agency (2025) The Seismological Bulletin of Japan. https://www.data.jma.go.jp/svd/eqev/data/bulletin/index_e.html. Accessed 1 Sep 2025\u003c/li\u003e\n \u003cli\u003eKato A (2024) Implications of fault-valve behavior from immediate aftershocks following the 2023 Mj6.5 earthquake beneath the Noto Peninsula, central Japan. Geophys Res Lett 51:e2023GL106444. https://doi.org/10.1029/2023GL106444\u003c/li\u003e\n \u003cli\u003eKatoh S, Nagao H, Imaizumi M (2024) Towards addressing challenges in seismic wave arrival-time picking models using deep learning. In: Proceedings of the Annual Conference of the Japanese Society for Artificial Intelligence. https://doi.org/10.11517/pjsai.JSAI2024.0_3L1OS3a02\u003c/li\u003e\n \u003cli\u003eKatoh S, Iio Y, Nagao H et al (2025) SegPhase: development of arrival time picking models for Japan\u0026rsquo;s seismic network using the hierarchical vision transformer. Earth Planets Space 77:118. https://doi.org/10.1186/s40623-025-02249-y\u003c/li\u003e\n \u003cli\u003eKhan S, Naseer M, Hayat M et al (2022) Transformers in vision: A survey. ACM Comput Surv 54:1\u0026ndash;41. https://doi.org/10.1145/3505244\u003c/li\u003e\n \u003cli\u003eKim A, Nakamura Y, Yukutake Y et al (2023) Development of a high-performance seismic phase picker using deep learning in the Hakone volcanic area. Earth Planets Space 75:85. https://doi.org/10.1186/s40623-023-01840-5\u003c/li\u003e\n \u003cli\u003eKubo H, Naoi M, Kano M (2024) Recent advances in earthquake seismology using machine learning. Earth Planets Space 76:36. https://doi.org/10.1186/s40623-024-01982-0\u003c/li\u003e\n \u003cli\u003eLiu S, Chen T, Chen X et al (2022a) More ConvNets in the 2020s: Scaling up kernels beyond 51\u0026times;51 using sparsity. https://arxiv.org/abs/2207.03620. [cs.CV]\u003c/li\u003e\n \u003cli\u003eLiu Z, Mao H, Wu CY et al (2022b) A ConvNet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, pp 11966\u0026ndash;11976. https://doi.org/10.1109/CVPR52688.2022.01167\u003c/li\u003e\n \u003cli\u003eM\u0026uuml;nchmeyer J, Woollam J, Rietbrock A et al (2022) Which picker fits my data? A quantitative evaluation of deep learning based seismic pickers. J Geophys Res. https://doi.org/10.1029/2021JB023499\u003c/li\u003e\n \u003cli\u003eNakagawa S, Tsuruoka H, Kato A, Sakai S, Hirata N (2016) A petabyte-scale large continuous seismic data analyzing system. Bull Earthq Res Inst 91:1\u0026ndash;9. https://doi.org/10.15083/0000032408 (in Japanese with English abstract)\u003c/li\u003e\n \u003cli\u003eNaoi M, Tamaribuchi K, Shimojo K et al (2024) Neural phase picker trained on the Japan Meteorological Agency unified earthquake catalog. Earth Planets Space 76:150. https://doi.org/10.1186/s40623-024-02091-8\u003c/li\u003e\n \u003cli\u003eNaoi M, Hirano S, Chen Y (2025) High-resolution monitoring of hydraulically induced acoustic emission activities using neural phase picking and matched filter analysis. Prog Earth Planet Sci 12:24. https://doi.org/10.1186/s40645-025-00696-5\u003c/li\u003e\n \u003cli\u003eNational Research Institute for Earth Science and Disaster Resilience (2019a) NIED Hi-net. https://doi.org/10.1759/NIED.0003\u003c/li\u003e\n \u003cli\u003eNational Research Institute for Earth Science and Disaster Resilience (2019b) NIED S-net. https://doi.org/10.17598/NIED.0007\u003c/li\u003e\n \u003cli\u003eNational Research Institute for Earth Science and Disaster Resilience (2019c) NIED DONET. https://doi.org/10.1759/NIED.0008\u003c/li\u003e\n \u003cli\u003eNational Research Institute for Earth Science and Disaster Resilience (2019d) NIED F-net. https://doi.org/10.17598/NIED.0005\u003c/li\u003e\n \u003cli\u003eRonneberger O, Fischer P, Brox T (2015) U-Net: Convolutional networks for biomedical image segmentation. In: Proceedings of the Medical Image Computing and Computer-Assisted Intervention. Springer, 9351:234\u0026ndash;241. https://doi.org/10.1007/978-3-319-24574-4_28\u003c/li\u003e\n \u003cli\u003eSandler M, Howard A, Zhu M et al (2018) MobileNetV2: Inverted residuals and linear bottlenecks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, pp 4510\u0026ndash;4520. https://doi.org/10.1109/CVPR.2018.00474\u003c/li\u003e\n \u003cli\u003eSifre L (2014) Rigid-motion scattering for image classification. Ph.D. thesis, \u0026Eacute;cole Polytechnique, Palaiseau, France.\u003c/li\u003e\n \u003cli\u003eSun H, Ross ZE, Zhu W, Azizzadenesheli K (2023) Phase neural operator for multi-station picking of seismic arrivals. Geophys Res Lett 50:e2023GL106434. https://doi.org/10.1029/2023GL106434\u003c/li\u003e\n \u003cli\u003eSuzuki R, Uchida N, Zhu W et al (2025) The forearc seismic belt: A fluid pathway constraining down-dip megathrust earthquake rupture. Science 389:190\u0026ndash;194. https://doi.org/10.1126/science.adt6389\u003c/li\u003e\n \u003cli\u003eTamaribuchi K (2018) Evaluation of automatic hypocenter determination in the JMA unified catalog. Earth Planets Space 70:1\u0026ndash;10. https://doi.org/10.1186/s40623-018-0915-4\u003c/li\u003e\n \u003cli\u003eTamaribuchi K, Moriwaki K, Ueno H, Tsukada S (2016) Automatic hypocenter determination for the Seismological Bulletin of Japan using Bayesian estimation. Quart J Seis 79:1\u0026ndash;13\u003c/li\u003e\n \u003cli\u003eTan YJ, Waldhauser F, Ellsworth WL et al (2021) Machine-learning-based high-resolution earthquake catalog reveals how complex fault structures were activated during the 2016\u0026ndash;2017 central Italy sequence. The Seismic Record 1:11\u0026ndash;19. https://doi.org/10.1785/0320210001\u003c/li\u003e\n \u003cli\u003eTan M, Le QV (2019) EfficientNet: Rethinking model scaling for convolutional neural network. In: Proceedings of the 36th International Conference on Machine Learning, pp 6105\u0026ndash;6114\u003c/li\u003e\n \u003cli\u003eWoollam J, M\u0026uuml;nchmeyer J, Tilmann F et al (2022) SeisBench\u0026mdash;a toolbox for machine learning in seismology. Seismol Res Lett 93:1695\u0026ndash;1709. https://doi.org/10.1785/0220210324\u003c/li\u003e\n \u003cli\u003eYu F, Koltun V (2016) Multi-scale context aggregation by dilated convolutions. In: Bengio Y, LeCun Y (eds) Proceedings of the 4th International Conference on Learning Representations, San Juan, Puerto Rico.\u003c/li\u003e\n \u003cli\u003eYu F, Koltun V, Funkhouser T (2017) Dilated residual networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, pp 636\u0026ndash;644. https://doi.org/10.1109/CVPR.2017.75\u003c/li\u003e\n \u003cli\u003eZhu W, Beroza GC (2019) PhaseNet: A deep-neural-network-based seismic arrival-time picking method. Geophys J Int 216:261\u0026ndash;273. https://doi.org/10.1093/gji/ggy423\u003c/li\u003e\n \u003cli\u003eZhu L, Peng Z, McClellan J et al (2019) Deep learning for seismic phase detection and picking in the aftershock zone of 2008 M7.9 Wenchuan earthquake. Phys Earth Planet Inter 293:106261. https://doi.org/10.1016/j.pepi.2019.05.004\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"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":"Neural phase picker, Deep learning, Hypocenter location, Earthquake catalog development","lastPublishedDoi":"10.21203/rs.3.rs-8174647/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8174647/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eTo develop a higher-quality seismic event catalog from Japan\u0026rsquo;s routine seismic observations, which have been continuously recorded at approximately 2000 stations, we trained deep-learning-based phase pickers using 20 years of arrival-time data read by the Japan Meteorological Agency (JMA). To enhance performance, we developed a new model, PhaseNeXt, by incorporating techniques proven effective in the field of computer vision\u0026mdash;particularly in semantic segmentation\u0026mdash;into PhaseNet, one of the most widely used neural phase pickers. The resulting model adopts an architecture inspired by DeepLab v3+, connecting parameter-efficient ConvNeXt blocks through residual connections, which mitigate vanishing gradients and allow scalable adaptation to larger training data sets. Furthermore, by including automatically read or briefly reviewed arrival-time readings that had not undergone detailed manual inspection in the training process, we demonstrated improved performance for small earthquakes. Using these insights, we trained three deep neural network models with different parameter sizes on 25\u0026nbsp;million waveforms associated with events listed in the 2002\u0026ndash;2023 JMA unified catalog. When integrated into the current JMA workflow, the best-performing model detected approximately 3.5 times more events than those listed in the JMA catalog while using nearly the same number of arrival-time readings.\u003c/p\u003e","manuscriptTitle":"PhaseNeXt: Neural phase picker trained on 20-year records to process the JMA-unified data set","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-22 16:38:42","doi":"10.21203/rs.3.rs-8174647/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Major Revision","date":"2026-03-23T21:01:40+00:00","index":"","fulltext":""},{"type":"reviewerAgreed","content":"","date":"2025-12-18T15:12:10+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-12-18T11:44:57+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-11-27T09:50:09+00:00","index":"","fulltext":""},{"type":"submitted","content":"Earth, Planets and Space","date":"2025-11-21T09:38:08+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":"aeb5d5ee-5e04-4a31-afe5-e276de9c3b98","owner":[],"postedDate":"December 22nd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-29T13:57:12+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-22 16:38:42","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8174647","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8174647","identity":"rs-8174647","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}