We recorded DAS data with a temporal sampling rate (i.e., ping rate) of 5000 Hz, which we decimated by a factor of 20 before storing them, yielding a sampling rate of 250 Hz. The system configuration has a gauge length of 10.2 m and a channel spacing of 2.04 m, resulting in a total of 7750 channels. Data is saved in 30 s segments (∼80 MB per file), resulting in a total size of 3.04 TB for the two-week experiment, with a short gap in data archiving occurring between 09:14 and 09:48 on 6 August 2024. The file name is UTC time +1 h (default naming from the system).

We performed tap tests on both Kefalonia and Ithaki islands to precisely identify the channel locations onshore. The cable onshore was visible along the entire length because it had been covered by asphalt along streets. Offshore, the cable path and channel positions are less certain. To reconstruct the offshore cable geometry and channel locations, we rely on the expected number of offshore channels (after subtracting those onshore) and refine the cable path by aligning the trend of the picked DAS P-wave arrival times with the earthquake-to-channel epicentral distances (using earthquake locations from NOA) (Fig. S1 in the Supplement). Automatic picking of P-wave arrivals along the cable result from applying PhaseNet-DAS (Zhu et al., 2023). We employ EMODnet bathymetry (https://emodnet.ec.europa.eu/en/, last access: 15 March 2025) to assign an elevation to the channels offshore, while for the onshore channels, we rely on the elevation from our GPS device. We also check for loops (typical of telecom infrastructure) in the onshore cable sections by analyzing the normalized energy recorded on each channel (Biondi et al., 2023b), and then cross-validate it with vehicle signals generated while driving along the cable (Fig. S2). We show the locations of identified loops in Fig. 2 (gray areas between dashed vertical lines).

We perform an initial assessment of data quality by manually inspecting recordings of earthquakes within a ∼50 km radius of the cable starting point using the revised NOA catalog (ML 0.5–3.4). The catalog lists 91 earthquakes within the two-week study period, all of which are clearly visible along the cable (Figs. 2, S3). Earthquake recordings show strong amplification (a factor >10–20) in the offshore segment, while onshore and offshore maximum noise amplitudes are similar, with slightly higher noise amplitudes offshore (Fig. 2). We observe amplitude scaling with earthquake magnitude that is consistent with theoretical expectations. For example, we obtain an amplitude ratio of ∼148 between an ML 0.6 and ML 3.0 event separated by ∼1.7 km (epicentral locations from the NOA catalog), as shown in Fig. 2c. We compute the amplitude for each event as a trimmed mean after excluding the upper and lower 10 % of values to reduce the influence of outliers. We band-pass filter the waveforms between 3–20 Hz and use channels 4000–5900. Assuming a theoretical scaling of 10ΔM, where ΔM is the magnitude difference between the two events, the observed amplitude ratio corresponds to an implied ΔM≈2.17. This value is slightly lower than the idealized expectation for perfectly co-located events and may reflect the minor differences in source location and measurement uncertainties. We note that magnitude estimate of the ML 0.6 event in the NOA catalog is based on two station observations. We also observe consistently elevated amplitudes at both the onshore–offshore transition zones of the cable (Fig. 2). In addition, we successfully record signals from teleseismic earthquakes, predominantly on the offshore portion of the cable (Fig. S4).

The seven seismic stations of the HUSN that complement the DAS data include: FSK (University of Patras, 2000), ATHR (University of Athens, 2008), VLS (National Observatory of Athens, Institute of Geodynamics, 1975), EVGI, DLMN, ITHC, DRAG (Aristotle University of Thessaloniki, 1981) (Figs. 1, S3).

We use a semblance-based earthquake detector, HECTOR, to detect earthquakes recorded along the fiber-optic cable (Porras et al., 2024). The detector evaluates the coherence of the waveforms along pre-computed hyperbolic trajectories with varying curvature and vertex along the cable and time axes (x–t) using a modified semblance function (Fig. 3). The run of the detector requires a set of pre-processing steps that include: (1) removing the mean and linear trend of each trace; (2) normalizing trace amplitudes to reduce the effect of geometrical spreading, cable coupling, and nonlinear effects; (3) applying a band-pass filter of 5–30 Hz as well as a (4) common mode noise removal via a frequency–wavenumber (fk) filter by removing the zero-wavenumber (k=0) component, which corresponds to coherent signals that arrive simultaneously across all DAS channels without exhibiting any phase delay.

Owing to the strong signal amplification offshore and the clearer hyperbolic seismic phase arrival moveout, we perform detection using an offshore segment of the cable (Fig. 2). Testing different configurations led us to use channels 4000–5900 for detection (Figs. 1c, 2), corresponding to a total length of ∼3900 m. Since the target seismicity cluster northwest of Kefalonia lies beyond the cable's extent, we fix the hyperbola vertex at channel 4000 (bottom Fig. 3c). We let the curvature of the hyperbola vary between 130–200° with steps of 5° and use temporal windows of 30 samples with steps of 10 samples to calculate the semblance and evaluate the coherence of the signals. We run the detection algorithm on 45 s data windows with a 15 s overlap (Fig. 3). The overlap ensures that potential earthquakes that might occur across two consecutive files are not missed. To eliminate duplicate entries, we de-cluster the initial detection list by removing signals that occur within 3 s of each other, retaining only the earliest detection. We choose this 3 s threshold specifically to account for the detector's inability to distinguish between P- and S-wave arrivals, which would otherwise register as separate events. Because the observed TS–TP times, where TS and TP denote S- and P-wave arrival times, along along the offshore channels do not exceed ∼2.5 s for earthquakes within cluster northwest of Kefalonia (Fig. 2a–b), a 3 s window safely merges these phases into a single event identification for most detections. The 15 s overlap between consecutive windows may also cause the same event to be detected twice, which we also mitigate with the de-clustering. We note that detection time (Fig. 3a) may be shifted with respect to the phase arrival observed in Fig. 3c because the vertex and the curvature of the hyperbola provide the highest signal coherence, which may not perfectly correspond with the earliest time of the phase arrival.

Running the HECTOR detection algorithm on the 2 weeks of data leads to a total of 6817 detections (Fig. 4). The number of detections increases substantially relative to the revised NOA catalog, with the highest counts occurring after ML 2.5+ earthquakes listed in that catalog (Fig. 4). The number of detections within 1 h bins following ML 2.5+ events from the NOA catalog exceeds 100 and exhibits an Omori-like decay (Fig. 4). We note that all ML 2.5+ NOA earthquakes reported in Fig. 4 occurred within the active cluster northwest of Kefalonia (pink box in Fig. 1). Importantly, although we calibrate the detector parameters to maximize sensitivity to the region of elevated seismicity offshore northwest of Kefalonia Island, we successfully detect all 91 earthquakes reported in the NOA catalog that occurred within ∼50 km of the cable's starting point (Fig. 1). The sensitivity to various azimuthal directions, distances, and magnitudes suggests that the set of parameters, DAS channels, and waveform processing used in the detector (as applied in this study) offer sufficient flexibility.

We manually inspect all detections and classify them into three groups based on their manually determined TS–TP values and arrival moveout: (i) local earthquakes (Figs. 4, 5a–b), (ii) distant earthquakes (with TS–TP>∼6 s) (Figs. 4, 5c), and (iii) false detections (Figs. 4, 5d–e). Local earthquakes (5734) include events within ∼50 km of the start of the cable and are identified by coherent signals along the nodes that commonly have clearly visible P- and/or S-wave arrivals (Fig. 5a–b). Smaller-amplitude earthquakes are recognizable primarily from characteristic reverberations between channels 5000–5500 in the submarine section of the cable (Fig. 5b). We inspect some of the weaker amplitude earthquakes with a spectral over-subtraction denoiser (Pascucci et al., 2026) to confirm that the signals correspond to real seismic events (Fig. S4). False detections (684) include non-seismic signals (e.g., boat noise; Fig. 5d), triggers within the coda of larger events where P- and S-wave arrivals are absent or unclear (Fig. 5e), and duplicate entries. In the case of duplicate detections, we retain the earliest detection (typically the P-wave arrival) and classify the latter detection as false. For example, we retain the P-wave arrival in Fig. 5c as the detection of the distant event, while the S-wave arrival is rejected as a false detection. We obtain a false detection rate of approximately 6 % (399 out of 6817 detections), which we consider a conservative overestimate. This is because some cases classified as false detections correspond to S-wave arrivals for which the associated P-wave is also detected, but the S-wave was not automatically removed during the 3 s de-clustering process. Manual inspection of signals also reveals missed events, particularly during the most active periods (i.e., after 8 August, Fig. S6).

We initially locate and calculate local magnitudes for high-SNR detected local earthquakes by combining DAS and seismic station data. We calculate SNR on offshore cable channels (4000–6800) considering a noise window from 1–5 s before the detection time and a signal window from 0.5–4.5 s after the detection time (Fig. 5). We compute the average Root-Mean-Square (RMS) values for noise and signal on traces band-pass filtered between 5–20 Hz, using a trimmed mean that removes the highest and lowest 10 % of values (trim fraction 0.1).

We then calculate SNR as follows: 1SNR=20log10(RMSsignal/RMSnoise).

We consider local earthquakes (following the criteria outlined above) with a SNR >12 dB for hypocentral location and magnitude estimation, resulting in a subset of 456 events out of the 5734 detected local earthquakes. We manually include a ML 3.4 earthquake on 8 August 2024 at 19:18 (event id: kef2196) despite its low SNR of 5.3 dB, as it represents the largest earthquake recorded northwest of Kefalonia during our study period. The low SNR results from the origin time being ∼12 s after a nearby ML 3.0 (event id: kef2195) whose coda wave contaminates the window containing the primary phase arrival of the ML 3.4 earthquake. Thus, our initial dataset consists of 457 earthquakes for location. The restrictive SNR threshold reflects the goal of building a high-quality reference catalog of locations and magnitudes. We perform automatic P- and S-wave arrival picking on DAS data using PhaseNet-DAS (Zhu et al., 2023) for each of the 457 earthquakes. We define 20 s windows for PhaseNet-DAS that start 6 s before the detection time and end 14 s after it (same as in Fig. 5). We follow the preprocessing used in the PhaseNet-DAS implementation (https://ai4eps.github.io/EQNet/phasenet_das/, last access: 26 May 2026), including conversion of strain to strain rate via temporal differentiation. We apply an additional band-pass filter between 5–40 Hz as part of our preprocessing. We manually inspect all picking results and observe good quality P- and S-phase picks (Fig. S7).

To reduce the number of DAS channels for earthquake location and achieve a number comparable to the seismic stations, we subdivide the cable into 12 segments (Fig. 1c) using a custom geometric clustering algorithm. The algorithm uses Density-Based Spatial Clustering of Applications with Noise (DBSCAN, Ester et al., 1996) to group channels exclusively by their spatial coordinates. It automatically instantiates a new segment boundary whenever the azimuthal angle variation exceeds a specified threshold. We allow a maximum azimuthal variation of 60°, applied an azimuth smoothing window of 50 channels, and enforce a minimum constraint of 100 channels per segment. We assign median P- and S-wave arrival times that we estimate in a prior step with PhaseNet-DAS applied to the entire cable to the median channel for each of the 12 segments. Finally, we download waveforms from the seven stations closest to the cable (Figs. 1, S3) and manually pick P- and S-wave arrival for each of the 457 earthquakes in Snuffler (Heimann et al., 2017) when data quality permits.

For clarity in our catalog enhancement workflow, we define template events as earthquakes with absolute hypocentral locations and magnitude estimates derived from the seismic network, and target events denote local earthquakes whose locations and magnitudes are initially unknown. We begin catalog enhancement by using 352 of the 356 located earthquakes as templates. We omit four events from the group of templates that have either a semi-major horizontal error ellipse axis and/or a depth error >5 km, or a depth above the sea level. This error threshold is necessary because large location uncertainties propagate into significant errors in ML, which in turn, affect the relative magnitudes that are calculated during enhancement using the amplitude ratios between templates and target events. Prior to calculating these cross-correlations, the continuous waveform records are downsampled to 50 Hz and the analysis is restricted to the same offshore cable section (channels 4000–5900) used for event detection (Fig. 7). We band-pass filter all traces between 5–20 Hz before computing a channel-based cross-correlation coefficient (CCC) using 4 s time windows that begin 0.5 s after the detection time (Figs. 7 and 8). We cross-correlate each of the 5378 target events with the 352 template events. Each target event is subsequently assigned the location of its most highly correlated template event. We retain the initial absolute locations and magnitude estimates in the final catalog for the four events previously excluded from the template database due to large hypocentral errors (>5 km) or because located above sea level.

We observe that CCC varies across channels (Figs. 7a, S10a) and that higher CCC values are associated with channels with higher SNR (Figs. 7c–d, S10c–d). To account for varying CCC with SNR, we apply a SNR criterion before averaging the CCC values of individual channels. We then compute the average CCC between event pairs using the Fisher mean (Fisher, 1921; Silver and Dunlap, 1987), after excluding 67 % of the channels (i.e., 1273 channels) with the lowest SNR (Figs. 7e, S10e). The SNR estimation uses the 4 s signal and noise windows noted above (Fig. 5), and the cutoff threshold is determined based on the lowest-amplitude earthquake in each event pair (Fig. 7b).

Before selecting the minimum CCC value for associating a target event with a template and proceeding with relative magnitude calculations, we first examine the expected decay of CCC with inter-event distance (Menke et al., 1990). We do so by cross correlating all independent template event pairs (excluding the four events noted above with larger hypocentral uncertainties) following the procedure outlined in the previous paragraph. We restrict the analysis to template earthquakes since we know their locations. Because our analysis relies solely on absolute event locations, the calculated inter-event distances for a given CCC value are likely to overestimate the true physical separation of the event pairs. As expected, we observe a systematic inverse correlation between CCC and inter-event distance (Fig. S11). For instance, we obtain a median inter-event distance of <2 km and a 95th percentile of ∼5 km for a pair with a CCC of 0.52 (Fig. S11). At 15 km distance, this corresponds to a magnitude error of 0.06 for a 2 km inter-event separation and 0.15 for a 5 km inter-event separation. At 10 km distance, the corresponding errors are 0.05 and 0.2, respectively. In this study, we use a CCC threshold of 0.52 that is empirically determined from the observed CCC–distance relationship. Specifically, Fig. S11 indicates a notable increase in inter-event distance for CCC values below ∼ 0.52–0.53. While median inter-event distances remain below ∼3 km for CCC values between 0.40 and 0.52, the 95th percentile increases substantially, indicating a higher risk of large spatial mismatches. We therefore exclude lower CCC values to ensure robustness of the event associations. By using a CCC threshold of 0.52, we associate 2515 target events to template events. We test slightly different CCC thresholds, namely 0.50 and 0.54, in addition to the preferred value of 0.52 used to construct the final enhanced catalog. We find that differences in the number of detected events occur primarily for magnitudes <-0.5 (Fig. S12). We provide all event-pair CCC values to enable reproducibility and allow assessment of the sensitivity of event associations to different CCC thresholds (see Data and code availability). Figure 8 shows examples of target-template pairs exceeding the CCC threshold. We observe that nearly all target events (>97%) are associated with template events from the cluster northwest of Kefalonia (Fig. S13).

Before calculating amplitude ratios for the determination of relative magnitudes, we first inspect whether we can correctly retrieve the magnitudes of template events using amplitude ratios, where template ML estimates originate from seismic stations (Fig. 9). We associate each template event with its most similar template and recalculate the magnitude from the amplitude ratio of band-passed traces (3–20 Hz) on channels 4000–5900. The lower corner frequency of 3 Hz suppresses long-period noise present in the offshore DAS segment. The upper corner frequency of 20 Hz is sufficiently high to capture the peak amplitudes of smaller earthquakes (ML≈0) with expected peak spectral amplitudes below the source corner frequency (e.g., Baltay and Abercrombie, 2025), and is sufficiently low to minimize high-frequency effects related to source complexity (e.g., Trugman et al., 2021). We consider the average of the maximum amplitudes of traces between 4000–5900 after removing the highest and lowest 10 % values (i.e., trimmed mean) for event pairs with CCC values >0.52. We calculate relative magnitudes as follows:

2Mtarget=Mtemplate+log10(Atarget/Atemplate) where M is the magnitude and A is the average of the maximum amplitude.

We observe good agreement between initial ML and relative ML estimates over the full magnitude range, as indicated by a coefficient of determination R2=0.92, mean absolute deviations of 0.1±0.1, and a best-fitting line with an approximately unit slope (Fig. 9). Remaining deviations between seismic-station ML and DAS-derived relative ML estimates may be attributed to imperfectly collocated event pairs and/or uncertainties in the initial magnitudes estimated from the seismic stations. Accordingly, we compute relative magnitudes for all target events with a minimum CCC >0.52 (2515 events) with respect to their template events.

Adding relative magnitude estimates to the 356 events with ML calculated from seismic stations leads to 2871 magnitude estimates (50 % of the local earthquakes). We observe that 2790 earthquakes with a magnitude estimate (97 %) are located offshore northwest of Kefalonia and are mostly clustered in space (Figs. 1, 5, S13). The remaining events are scattered within the study region (Fig. S13), with a few clusters located along the Strait of Ithaki. Figure S13 shows the distribution of all events detected and located in this study with a magnitude estimate. There are 10 events in common with the revised NOA catalog (the NOA catalog reports 91 earthquakes within a radius of 50 km from the start of the cable) that are not included in our catalog of 356 events, despite being successfully detected. Eight of the ten events common to the NOA catalog are omitted because of their average SNR at offshore DAS channels being <12 dB, and they were therefore not considered for initial location (Sect. 4.1). The last two of the ten events are omitted because we observe TS–TP>6 s at offshore DAS channels. The fact that we detect but do not locate a few earthquakes included in the revised NOA catalog reflects our focus on detecting and locating seismicity in the highly active region offshore northwest of Kefalonia. By considering events within the active region northwest of Kefalonia (Figs. 1, 5, S13), we observe a ∼38-fold increase in the number of events with a magnitude estimate relative to the revised NOA catalog (Fig. 10). All events in our catalog down to ML 2.4 are common to the NOA catalog (Fig. 10a), while the number of events from this study significantly increases at lower magnitudes (Fig. 10). Specifically, we estimate a magnitude of completeness of Mc=-0.4 using both a maximum curvature and Goodness-of-Fit Test (Wiemer and Wyss, 2000) and a b-value =0.95±0.03 based on the maximum likelihood estimate (Aki, 1965) (Fig. 10a). The enhanced earthquake catalog shows inter-event times as short as 3 s (the minimum allowed inter-event time from de-clustering) and average inter-event times for 100-event windows down to 30 s.

This study releases two-week DAS waveforms and a high-resolution earthquake catalog derived from the integration of DAS recordings and publicly available data from conventional seismic stations. The dataset captures a period of exceptionally high seismicity within 10–15 km of the fiber-optic cable at rates exceeding 100 events per hour (Figs. 1, 6). Although other open-access DAS datasets also contain high event rates, most of them were acquired in boreholes and target induced seismicity (the FORGE dataset, Porras et al., 2024), rather than natural earthquake sequences.

The highest seismicity rates occur offshore in a cluster northwest of Kefalonia with a ∼5 km radius where azimuthal coverage from land stations is limited (Figs. 1, 6). The intense seismic activity combined with sparse station geometry are challenging for the monitoring of active seismic sequences (Grigoli et al., 2021; Karastathis et al., 2015). This work shows that combining DAS and seismic station data can make it possible to overcome some of the challenges associated with observing offshore earthquakes: using DAS channels for dense wavefield sampling and earthquake detection, combining DAS and seismic station data for determining hypocenter locations, and using seismic stations for amplitude calibration and magnitude estimation.

The enhanced catalog presented here exhibits high temporal resolution, with inter-event times as short as 3 s and a completeness magnitude of Mc=-0.4. It captures evolving statistical features of the seismic sequence, including a transition from more Poisson-like inter-event time distribution to densely clustered activity associated with mainshock–aftershock activity (Fig. 10c). The catalog also reveals notable changes in aftershock productivity. For instance, two M>3 earthquakes before 8 August 2024 were not followed by clear aftershock activity, whereas similar-sized and smaller events after that date produced pronounced aftershock sequences (Fig. 10b–c). After 8 August, we observe a systematic relation between mainshock magnitude and aftershock inter-event times, with larger mainshocks followed by shorter aftershock inter-event times (Fig. 10d). The distribution of the seismicity highlights a nearly vertical WNW-ESE striking fault (Fig. 6), which is consistent with the geometry observed by Anagnostou et al. (2025), who used 3-month data between the end of February and the end of April 2025. These observations underscore the high spatial and temporal resolution of the catalog that we will discuss in a detailed analysis of the space–time–magnitude patterns that will be addressed in a separate study.

The amphibious nature of the telecommunication cable introduces additional features of shallow crustal structure that warrant further investigation, including the strong signal amplification observed along the marine segment and its potential influence on magnitude determination when using DAS data alone. This offshore amplification may reflect not only variations in shallow crustal structure, but also differences in DAS sensitivity to body and surface waves, mechanical coupling, cable construction, and installation conditions across the onshore–offshore transition.

The datasets released in this work may offer a range of potential applications. The exceptionally high seismicity rate with inter-event times as short as 3 s provides an ideal testbed for developing and evaluating advanced detection and location methods tailored to DAS data. To support such efforts, we provide a manually verified event catalog that can serve as a reference dataset. In addition, the inclusion of very small-amplitude signals, some with magnitude estimates and others without, may be useful for developing and testing denoising approaches (e.g., Fig. S5).

The integration of the DAS array within an open-access seismic network offers an opportunity to evaluate DAS-enhanced event location methods and explore hybrid location strategies. This setting also allows testing of various strategies to mitigate biases that could arise from the large differences in the number of sensors between the DAS cable and the traditional network, such as automatic DAS channel selection. Furthermore, DAS recordings acquired on seafloor cables may exhibit higher amplitudes compared to onshore cables, as observed in this study. Such differences could reflect a combination of site, coupling, and propagation effects that may in turn introduce biases in magnitude estimation. Finally, the dataset may offer valuable opportunities for machine learning applications. With nearly 6000 recorded events, it can be used to train new models or to investigate transfer learning of existing ones.

Continuous DAS data waveforms are available from Bocchini et al. (2025), 10.60517/cv43p1601.

Codes to read and plot the events (or detections) recorded along the fiber-optic cable, earthquake catalogs, event pairs cross-correlation coefficients, cable geometry, and PhaseNet-DAS picks are available from Bocchini (2026), 10.5281/zenodo.20558686.

We use _PhaseNet-DAS (Zhu et al., 2023) for phase picking on DAS data, which is available from https://ai4eps.github.io/EQNet/phasenet_das/ (last access: 26 May 2026).

We use HECTOR (Porras et al., 2024) for event detection on DAS data, which is available from https://github.com/juanucr/HECTOR/tree/main (Porras, 2024).

We use Snuffler, available through Pyrocko, for manual picking of seismic waveforms (10.5880/GFZ.2.1.2017.001, Heimann et al., 2017).

We use Obspy for processing of seismic station waveforms (Beyreuther et al., 2010).

We create figures using Matplotlib (Hunter, 2007) and GMT (Wessel et al., 2019).

In this study, we combine DAS recordings with permanent seismic stations from the HUSN to investigate seismicity on Kefalonia Island between 2–15 August 2024. Applying a semblance-based detector enables identification of more than 5700 earthquakes within a ∼50 km radius of the starting point of the fiber-optic cable. Combining DAS and seismic data analysis and waveform cross-correlation enables location and magnitude estimates of 2871 events, of which 2790 concentrate in an active cluster offshore northwest of Kefalonia. The catalog shows very high seismicity rates with peaks in earthquake rates that exceed 100 events per hour. Individual inter-event times are as short as 3 s, while average inter-event times calculated over moving 100-event windows decrease to ∼30 s. The short inter-event times and a Mc of -0.4 enable a detailed spatiotemporal analysis of seismicity.

A comparison of the catalog produced here and the revised earthquake catalog of NOA for the active cluster northwest of Kefalonia shows the same number of events down to ML 2.4, and a factor of ∼38 increase in smaller events for the DAS-seismic-data catalog. The detailed catalog underscores the benefit of using DAS to strengthen seismic monitoring in regions where station coverage is limited and seismicity is intense.

This work makes the following data products publicly available: the DAS-seismic station earthquake catalog, continuous DAS recordings, detection lists, and supporting metadata. The dataset provides a resource for future studies of DAS data processing, integration with seismic stations for earthquake monitoring, and analyses of seismicity, in large part because it targets a highly productive natural seismic sequence.