A Distributed Acoustic Sensing Dataset for Vessel Detection and Localization in Submarine Cable Protection Erick Eduardo Ramirez-Torres1 , Javier Macias-Guarasa1,5 , Daniel Pizarro1 , Javier Tejedor2 , Sira Elena Palazuelos-Cagigas1 , Pedro J. Vidal-Moreno1 , Marı́a R. Fernández-Ruiz1 , Sonia Martin-Lopez1,3 , Miguel Gonzalez-Herraez1,3 , and Roel Vanthillo4 1
arXiv:2607.28306v1 [physics.geo-ph] 30 Jul 2026
2
Universidad de Alcalá, Departamento de Electrónica, Alcalá de Henares, Spain Institute of Technology, Universidad San Pablo-CEU, CEU Universities, Urbanización Monteprı́ncipe, 28668 Boadilla del Monte, Spain 3 Daza de Valdés Institute of Optics (IO-CSIC), Madrid, Spain 4 Marlinks, Leuven, Belgium 4 Corresponding author
Abstract Recent incidents of accidental damage and suspected sabotage to submarine telecommunication and power cables, particularly in the Baltic Sea, have underscored their vulnerability and the need for continuous monitoring solutions. Distributed acoustic sensing (DAS) applied to submarine optical-fiber cables enables wide-area monitoring of underwater acoustic activity. We present the Marlinks-NS DAS dataset, comprising processed submarine DAS measurements and AIS-derived vessel information curated for cable-protection research. The dataset defines two machine-learning tasks (vessel detection and vessel-to-cable distance estimation) allowing reproducible research under realistic marine conditions. The dataset contains 74,771 labeled data instances from ten days of continuous recording along a 2, 554 m segment in a 28 km buried fiber-optic cable in the North Sea. Each instance includes spectral-energy features from 250 sensing channels, together with anonymized distance measurements and metadata from AIS information. The released HDF5 data, documentation, processing description, and example code support reproducible development and evaluation of DAS-based vessel-monitoring methods for submarine cable protection.
1
Background & Summary
Submarine cables are vital for global communication and power transmission but remain vulnerable to accidental damage and sabotage, with recent severe examples in the Baltic Sea. These include, for example, the damage to the EstLink2 power cable [1], and the C-Lion1 communication cable [2], both suspected of sabotage and linked to geopolitical tensions. Distributed Acoustic Sensing (DAS) technology allows repurposing of standard optical fibers into dense acoustic sensor arrays capable of capturing vibrations and strain over tens of kilometers. In the context of submarine cables, this capability enables continuous monitoring of marine activity using existing or dedicated fiber-optic infrastructure. Unlike radar or satellite-based systems, DAS provides continuous, real-time coverage, is largely unaffected by lighting or weather conditions, and does not depend on cooperative identification systems such as the Automatic Identification System (AIS). Publicly available datasets enabling the development of data-driven algorithms for DAS-based vessel detection are scarce. Existing data resources are often limited in scale, or lack properly labeled metadata that allow research in Artificial Intelligence/Machine Learning (AI/ML) strategies. This data scarcity is particularly severe in specialized environments like submarine cables. Initiatives such as PubDAS [3] (the first large-scale open-source repository for DAS data) have allowed accelerated 1
Table 1: Summary of publicly released DAS datasets cited in this work, including application domain, fiber length, dataset size, sampling frequency, gauge length and annotation availability. Dataset
Application domain Fiber length
DAShip [4, 5]
8.5 km Wake ship detection and vessel speed/angle classification
MARS DAS exper- Ocean dynamics & iment [6, 7] subsurface characterization MEUST [8, 9] Seismic monitoring
Dataset size Sampling frequency and gauge length (GL) 10 Hz 55875 ship GL=4 m passing events above the cable 4 days, 3.2TB 250 Hz GL=10 m
51 km
41.5 km
6 days 1 day (68 GB), 7 days (740 GB), 20 days (16 TB) 44 days
HCMR+ NESTOR+ MEUST [10, 11]
Seismic monitoring
13.2 km, 26.2 km and 44.8 km
DAS4Microseism [12, 13]
Ocean dynamics & subsurface characterization Seismic monitoring
120 km
OOI RCA [14, 15] Valencia [3]
Yellow River Delta [16, 17]
95 km
Submarine monitoring ∼50 km (first vibrations (telecommu- ∼10 km on nication cable) land) Nearshore ocean cur10 km rent monitoring
Annotation
Annotated using AIS guidance (for the three binary classification tasks) No labeling is provided
1 kHz No labeling is provided GL=19.2 m No labeling is provided Between 100 Hz and 500 Hz. GL=19.2 m/10 m 50 Hz No labeling is provided GL=8-16 m
4 days, 2.4TB 200 Hz No labeling is provided GL=40.852 m 7 days, 3TB 250 Hz No labeling is provided GL=30.4 m 25 days
50 Hz GL=5 m
Provide labels for wind speed and tidal data
research by providing multiple experiment datasets for benchmarking and algorithm development. However, the number of publicly available resources in submarine environments is very limited. Some relevant examples in this scenario are: • The DAShip dataset, by Huang et al. [4] (available at [5]). • The MARS DAS experiment dataset by Cheng et al. [6] (available at [7] with a repository at GitHub). • The MEUST DAS dataset by Sladen et al. [8] (available at [9]). • The HCMR+NESTOR+MEUST DAS datasets by Lior et al. [10] (available at [11]). • The DAS4Microseism dataset, by Taweesintananon et al. [12] (available at [13]). • The OOI Regional Cabled Array (RCA) Dataset, by Lipowsky et al. [14] (available at [15] with a source code repository at GitHub). • The Valencia-IslaLink DAS dataset, by Spica and Gaite, which is distributed within PubDAS [3]. • The Yellow River Delta Dataset, by Song et al. [16] (available at [17]). Table 1 shows the relevant characteristics of each of them. Except for DAShip, all these datasets are oriented to applications in the seismic and ocean monitoring domain without explicit labeling information that can support experimental work in vessel detection and localization. DAShip is a large-scale, annotated dataset collected for the specific purpose of marine vessel detection. It comprises 55,875 ship-related event segments, captured on a submarine fiber-optic cable, 8.5 km 2
long, with a maximum depth of 13 m along the cable deployment (according to GEBCO bathymetry information [18]). It is oriented towards detecting vessel passages directly above the fiber-optic cable using imageclassification algorithms applied to wake signatures. This formulation is less suited to early-warning cable-protection scenarios, in which vessels should ideally be detected before reaching the cable. It is oriented towards detecting vessel passages right above the fiber-optic cable, by using image classification algorithms to identify the ships’ wakes, so that This formulation is less suited to early-warning cable-protection scenarios, in which vessels should be detected well before reaching the cable. Another limitation is that the data have been downsampled to 10 Hz, which has two main issues: first, this reduced bandwidth limits the possibility of exploiting the higher frequency components known to be significant in vessel acoustic signatures [19, 20]; and second, it will be more affected by low-frequency noise from temperature drift and the effect of sea waves and marine currents. We can mention other properly labeled recent datasets which address event classification tasks, but in terrestrial domains, such as that by Tomasov et al. [21] that presented a fully labeled DAS dataset, including band-limited feature vectors. The fiber was 1,663 m long, and was buried 1 m below the surface around a rectangular area on an university campus. The selected events included walking, running, longboarding, and driving, as well as potential security-related events like fence climbing, fiber manipulation, and opening/closing of manholes. The Marlinks-NS DAS dataset that we present in this paper addresses the lack of publicly accessible, well-annotated DAS data aimed at submarine cable protection applications, by defining two AI/ML tasks, one for vessel detection, and another for vessel distance estimation. The dataset provides a large-scale, machine-learning-ready collection of preprocessed spectral features and vessel metadata derived from a ten-day continuous recording campaign on a submarine cable. It comprises 74,771 data instances (feature vectors) computed at 250 channels in a relevant fiber-optic sensing range, each containing energy values in 100 logarithmically-spaced frequency bands, together with synchronized vessel distance labels, plus anonymized vessel-related information obtained from AIS (type, beam, and length). By releasing this curated dataset, we aim to promote transparency and reproducibility in AI/ML-based DAS research while enabling the scientific community to benchmark vessel detection and localization algorithms under realistic conditions. Basic data-handling examples are archived with the dataset on Zenodo, while additional actively maintained reproducibility tools are available in the companion GitHub repository at https://github.com/UAH-PSI/das-vessel-detection.
Figure 1: Dataset curation and processing pipeline, covering DAS acquisition and raw signal preprocessing; spectral feature extraction; metadata (AIS, geographical location and bathymetry) processing, synchronization and label generation; and cross-validation partitioning.
3
The DAS and AIS measurements underlying this dataset were previously analyzed in [22] to develop and evaluate machine-learning methods for vessel detection and vessel-to-cable distance estimation. That research article focused on algorithm design and application performance, including spatial and temporal aggregation, model comparisons, beamforming-based localization, vesseldependent error analysis, and computational requirements. The present Data Descriptor provides dataset-centered documentation and reproducibility material that complements that applicationcentered study, rather than repeating its broader experimental analysis. Specifically, it provides a more detailed account of dataset generation and curation, the released feature representation, HDF5 organization, repository contents, metadata, known limitations and archived data-handling code, together with a new passage-level characterization of maritime-traffic diversity across the recommended day-wise folds, an expanded characterization of environmental variability, and baseline technical validation to support independent reuse.
2
Methods
The processing pipeline we applied for the generation of the dataset is shown in Fig. 1, to provide a full reference about the data curation process. All the relevant methods and modules are described in the following subsections.
2.1
DAS interrogator: Acquisition Method and Setup
The submarine fiber-optic cable was interrogated using a phase-sensitive Optical Time Domain Reflectometry (ϕ-OTDR) system. This approach uses a narrow-linewidth, highly-coherent laser that emits short optical pulses into the fiber core, where microscopic refractive index variations act as Rayleigh scatterers. The backscattered signal from each pulse carries amplitude and phase information that varies with longitudinal deformation in the fiber, which can be related to vibrations and acoustic disturbances in the cable vicinity. By measuring phase differences between consecutive backscatter traces, dynamic strain variations can be reconstructed continuously along the entire length of the fiber. The DAS data were recorded during a ten-day acquisition campaign from 16 to 25 June 2023 using an Alcatel OptoDAS interrogator [23] connected to a 28 km ocean-bottom fiber-optic cable originally deployed for power cable monitoring. The cable lies buried between 1.4 − 7.2 m below the seafloor, offshore Zeebrugge (Belgium) in the North Sea, as shown in Fig. 2. The interrogator is located off-shore and employs optical pulse-compression reflectometry, which enables distributed vibration sensing over tens of kilometers with meter-scale spatial sampling, while mitigating fading effects due to low-intensity points in the Rayleigh pattern [24, 25, 26]. It has a gauge length of L = 10.21 m, hence generating 2,774 raw spatial channels. The raw differential phase-rate signals were sampled at fs = 3,125 Hz.
2.2
DAS Preprocessing
In this stage, the raw differential phase-rate time series generated by the DAS interrogator are converted to strain. This procedure is very closely related to the interrogator design and characteristics, so that the corresponding modules are mainly provided by the cable operator. The operations carried out are: • Raw-signal scaling: The integer samples delivered by the OptoDAS system were mapped to differential phase rate using the instrument scaling factor. Processing was performed separately for each channel along the time axis. • Impulse suppression: Before integration, isolated spikes were identified using a channeldependent local criterion. The threshold combined a 50-sample running mean with an offset of 30 rad/(m s), and samples exceeding it in magnitude were replaced by zero. These parameters were selected empirically from separate DAS recordings outside the released campaign. 4
(a) Regional map.
(b) Local map showing cable (red line).
(c) Local map showing cable (red line) and bathymetry.
Figure 2: General location and bathymetry (cable location has been displaced for security considerations), from [22]. • Temporal phase unwrapping: Each channel was unwrapped independently in time to remove discontinuities of 2π. • Strain recovery: The cleaned and unwrapped phase-rate sequences were integrated over time and converted into strain, expressed in m/m. Figs. 6.a+b and 5.a+b show two examples of the generated strain signals for scenarios with a vessel nearby and in the absence of nearby vessels, respectively.
2.3
Metadata Processing: AIS Data Processing
The AIS data covering the same acquisition period, provided by the cable operator, included vessel identifiers (MMSI), timestamps, coordinates, course, and speed for 745 unique vessels with 64,417 position reports. A relevant issue when using AIS data is the unpredictable and typically low position update rates [27]. A detailed analysis on our DAS data showed that 88% of vessels reported positions only every 1 to 3 minutes, which may imply a significant positional uncertainty. To overcome this issue, we applied a linear interpolation between consecutive reported positions at 1 Hz position update rate, excluding periods in which vessels did not report for time intervals over 60 minutes. Because the AIS records were the only available source of vessel positioning, the accuracy of the interpolated trajectories could not be assessed against an independent reference. Linear interpolation was retained as a transparent baseline method. The resulting distance labels therefore remain subject to uncertainty arising from the AIS reporting interval and from vessel motion between consecutive reports. The application-based results reported in Section 4 indicate that these labels retain information relevant to the defined ML tasks, but they do not constitute a direct validation of interpolation accuracy. More sophisticated trajectory-reconstruction methods could be investigated in future work [28]. After the interpolation process, we had over 1, 200, 000 AIS position entries, which we further enriched with web-scraped additional AIS metadata (namely, vessel type, length and beam) that can be useful in further studies or AI/ML tasks requiring information related to vessel type and size.
2.4
Metadata Processing: Geographical+Bathymetry Data Processing
In our DAS context, geospatial tagging refers to the process of assigning precise geographical coordinates, and depth or elevation to each sensed position along the fiber-optic cable. In DAS
5
applications, accurate geospatial tagging is crucial because it relates the distributed acoustic measurements to their exact physical locations along the fiber. This enables spatially consistent signal interpretation, correlation with environmental and bathymetric data, and the development of position-dependent models for event localization and propagation analysis [29]. In our dataset, each DAS sensing position along the selected cable segment was assigned precise coordinates provided by the data owner, which resulted in a complete set of geospatial attributes (longitude, latitude, depth) for all the DAS channels, which provides a fundamental reference in our geographically-informed ML tasks using the Marlinks-NS DAS dataset. Cable and vessel horizontal coordinates were represented in the WGS 84 geographic coordinate system (EPSG:4326), using owner-supplied cable elevations and an assumed vessel elevation of 0 m; coordinates were transformed to WGS 84 Earth- centered, Earth-fixed coordinates (EPSG:4978) using EcefKarney from PyGeodesy 23.1.9, and vessel-to-cable distances were computed as the minimum three-dimensional Euclidean chord distance between each vessel position and the interpolated cable sensing points.
2.5
ML Tasks and Data Balance Considerations
In ML applications, maintaining a balanced dataset helps ensure unbiased model training and fair performance evaluation. For classification tasks, balance in the number of samples per class prevents the model from favoring dominant categories. For regression tasks, an even distribution of target values enables the model to learn the underlying functional relationships for the full data range, rather than overfitting to the most densely sampled regions. However, achieving such balance is often difficult in real-world scenarios, where natural or operational processes generate highly uneven data distributions, which is usually the case in AI/ML-based DAS applications. The Marlinks-NS DAS dataset has been designed to support two complementary ML tasks related to vessel monitoring for submarine cable protection. The first task, vessel detection, is formulated as a binary classification problem that is oriented to discriminate whether there is a vessel closer than a given distance threshold to the cable or not. The second task, distance estimation, is defined as a regression problem in which the model aims to predict the distance between the cable and the nearest vessel. In both cases, the required labeling information is derived from the synchronized AIS data.
Figure 3: Number of data frames corresponding to vessels located at different distance ranges from the cable. 6
To assess whether both tasks are representative, and evaluate their statistical balance, a detailed analysis of the vessel distance distribution along the monitored cable was carried out. Initial inspection of the ten-day continuous recording revealed strong non-uniform spatial distribution. Fig. 3 shows, as a function of the channel position along the cable length, the number of 10-second data frames that corresponds to a situation in which there is a vessel at different distance ranges. From this figure we can identify a cable section around 16 km (channel 1565, roughly in the middle of the vertical yellow transparent band in the Figure) that exhibits a relevant increase in the number of available data frames with vessels closer to the cable than, for example, 1,000 m or 2,000 m, which can be candidate thresholds for the vessel detection task. As an example of the rationale for defining a “reasonable threshold”, a Spanish company monitoring submarine cables with AIS, sets a 1,000 m distance threshold to trigger enhanced vessel tracking to assess potential threats to the cable. This region corresponds to the location of the cross between the submarine cable and a dredged fairway (approximately 20 m deep) providing harbor access, which is consistent with the higher density of vessels shown in Fig. 3. Therefore, on the basis of this spatial distribution, we selected a contiguous cable range of 2,553 m, corresponding to a subset of Nchannels = 250 equally-spaced spatial channels located between 14,702 m and 17,255 m from the interrogator. This segment exhibits a clear unbalance between the “vessel-nearby” and “no-vessel-nearby” classes (e.g. 22,347 vs. 52,424 for the 1,000 m distance threshold), but provides a practical compromise between the availability of vessel-nearby observations and the preservation of a contiguous sensing range. Finally, to give an idea on the distribution of vessel types, Fig. 4 shows the distribution boxplots of minimum distances by vessel types, in which the top numbers (n = . . . ) state the number of examples of each vessel type in the dataset.
Figure 4: Boxplot of data frames available as a function of the minimum distance of any vessel to the cable, per vessel type.
2.6
Feature Extraction
To generate the dataset, we adopted a data-driven methodology to decide on the feature vector composition, informed by the available AIS data. The main idea was to compare long-term averaged strain spectra corresponding to vessel and background-noise conditions and identify frequency regions containing potentially discriminative information. Fig. 5.c shows the long-term averaged strain spectra (in dB) for the case of vessels closer than 1,000 m to the cable (Vessels 1000m trace) and for the case of no vessels being closer than 3,000 m (Noise 3000m trace). Based on this analysis, the lower cutoff was conservatively set to 4 Hz to reduce the contribution of slow components associated with temperature variations, surface-wave loading, and other slowly varying environmental effects. Therefore, 4 Hz should be understood as a processing choice intended 7
Figure 5: (a) Filtered strain signal when no vessel is nearby the cable (it corresponds to the right dashed square in Fig. 6.c). (b) Filtered strain data as a function of time and distance corresponding to the right red dashed square in Fig. 6.c, time synchronized with Fig. 5.a. (c) Long term averaged strain power spectrum plots (expressed in dBs) for strain measurements for vessels closer than 1,000 m to the cable (Vessels 1000m trace), and no vessels closer than 3,000 m (Noise 3000m trace). to improve the robustness of the released representation, rather than as a precise physical boundary for vessel-generated signals. The upper cutoff was selected considering both the observed spectra and the spatial response of the DAS acquisition configuration. Assuming that the gauge length was equal to the channel spacing, L = ∆x = 10.21 m, and using the apparent propagation velocity estimated for the dominant propagation path, capp ≃ 1,750 m/s [22], the nominal along-fiber spatial-aliasing limit is falias = capp /(2∆x) ≃ 85.7 Hz. Independently, the spatial averaging over the gauge length progressively attenuates the DAS response, with its first theoretical notch occurring at fnotch = capp /L ≃ 171.4 Hz under the same propagation assumption.
Figure 6: (a) Filtered strain signal sensed at 15.4 km in the time interval of the left red dashed square in Fig. 6.c (4 Hz to 98 Hz, excluding the (49 − 51) Hz). (b) Filtered strain data as a function of time and distance along the cable corresponding to the left red dashed square in Fig. 6.c, time synchronized with Fig. 6.a. (c) Feature vector energy (in dBs) as a function of time and distance along the cable. (d) Distance from cable to nearest vessel (in meters), time-synchronized with Fig. 6.c. The spatial-aliasing limit does not constitute a sharp temporal-frequency cutoff: components above it can still be measured at individual channels, although their spatial variation may be 8
aliased, while gauge-length averaging increasingly attenuates them as the frequency approaches the first notch. Since the empirical spectra continued to show vessel-to-background differences close to 100 Hz, this frequency was retained as the upper limit of the initial analysis range, provided the very small amplitude of components at higher frequencies. Additionally, we discovered strong narrowband spectral components around 100 Hz, with harmonics at its integer multiples and a subharmonic around 50 Hz, probably due to interference generated by rotating mechanical systems or electrical noise in the interrogator environment. Since these components were consistently present under both vessel and background-noise conditions, they were considered non-discriminative site- or instrument-related noise, and the bands associated with the nominal 49–51 Hz and 98–100 Hz intervals were excluded from the feature representation. The strain signal for each channel was then segmented into non-overlapping windows of length Tω = 10 s, to which we apply a Blackman window to reduce spectral leakage. At the sampling frequency fs = 3,125 Hz, each window contains NFFT = 31,250 samples, resulting in a nominal FFT-bin spacing of ∆f = fs /NFFT = 1/Tω = 0.1 Hz. The window length was selected as a compromise between temporal granularity and frequency-domain representation: it provides sufficiently fine spectral sampling in the low-frequency region while generating one feature vector every 10 s, allowing the temporal evolution of vessel-induced signals to be reasonably preserved. The FFT coefficients within the initial 4–100 Hz analysis range were then retained for the subsequent bandenergy calculation, with the noise-related intervals excluded during the frequency-band definition described below. The feature vector is then generated from the FFT values, so that each 10-second window is represented by Nbands = 100 logarithmically spaced band-energy features. To construct these bands, we first generated N0 = 104 contiguous frequency intervals over the initial 4–100 Hz analysis range, using the logarithmically spaced edges described by fi = fmin
fmax fmin
i/N0 ,
i = 0, 1, . . . , N0 ,
(1)
where fmin = 4 Hz and fmax = 100 Hz. N0 = 104 is used so that the final number of frequency bands is Nbands = 100, given that bands overlapping the nominal 49–51 Hz interference interval were subsequently discarded, together with the highest-frequency band overlapping the nominal 98–100 Hz interval. This geometric construction provides a constant ratio between consecutive frequency edges and, therefore, an approximately constant fractional bandwidth. Consequently, the bands are narrower in absolute frequency at the lower end of the spectrum, providing a finer representation of the low-frequency region, and become progressively wider at higher frequencies. The exact limits of all the retained frequency bands are provided in the fbands.csv file, available in the archived Zenodo release and in the companion GitHub repository. The mathematical details of the feature extraction calculation are as follows. We assume that (n) xc [m] denotes the discrete version of the nth Blackman-windowed strain signal corresponding to channel c and starting at n · Tω seconds, with n = 0, 1, . . . , Nwindows − 1 and c = 0, 1, . . . , Nchannels − 1, where Nwindows and Nchannels denote the number of available temporal windows and sensing (n) channels, respectively. The sample index is m = 0, 1, . . . , NFFT − 1. We then define Xc [k] as the (n) one-sided NFFT -point FFT of xc [m], with NFFT = 31,250 and k = 0, 1, . . . , NFFT/2. Each spectral band b = 0, 1, . . . , Nbands − 1 is defined by its FFT-bin limits kb− , kb+ , where 0 ≤ kb− ≤ kb+ < NFFT (corresponding to the frequency band limits defined in Eq. (1)). From these definitions, the spectral energy of band b for window n of channel c is computed as: +
Ec(n) [b] =
kb X
2
Xc(n) [k] .
(2)
k=kb−
h i (n) (n) (n) (n) (n) Finally, the feature vector for xc [m] will be Ec = Ec [0], Ec [1], . . . , Ec [Nbands − 1] . For each temporal window, the feature vectors from all channels form an Nchannels × Nbands feature 9
matrix. Stacking these matrices over time produces a tensor of dimensions Nwindows × Nchannels × Nbands , which is stored in the X dataset of the released HDF5 file.
2.7
Data Synchronization and Labeling
Time synchronization between the DAS signals and the interpolated AIS information is carried out as a required step before data labeling is carried out. The interpolated AIS positions are then used to continuously compute the shortest distance between each vessel and every DAS sensing point. These distance values are considered within each 10-second strain signal window to determine the minimum vessel-to-cable distance used as the primary label for the ML tasks. This synchronization process ensures that all DAS feature vectors are associated with consistent, temporally-aligned vessel metadata. The closest distance continuous variable supports regression tasks directly as Vessel distance labels, or can be converted into vessel detection labels for the classification task, by simply applying user-defined distance thresholds.
3
Data Records
The dataset and its accompanying resources are publicly available on Zenodo [30]. The primary data product is the approximately 14-GB HDF5 file dataset sensor range 1440 1690 0.h5. The Zenodo deposit also includes dataset documentation in Markdown and PDF formats (README.md and README.pdf), supporting metadata in misc.zip, and simple Python data-handling examples in src.zip. The structure of the primary HDF5 data file is summarized in Table 2 and described below: • X: A 3D NumPy array of shape (Nwindows , Nchannels , Nbands ), containing the feature vectors in squared strain units (m/m)2 , captured by the sensors at each timestamp. – Nwindows = 74,771: number of non-overlapped 10-second signal windows analyzed along the full recording period. – Nchannels = 250: number of spatial channels (sensor positions along the selected fiber segment). – Nbands = 100: number of energy-band features per channel. • y: A 1D NumPy array of shape (Nwindows ) containing the distance (in meters) to the closest vessel during each 10-second window. This continuous variable supports regression tasks directly, or can be converted into classification labels by simply applying user-defined distance thresholds. • datetimes: An array of strings in the HDF5 file that records the UTC timestamp (formatted as %Y-%m-%d %H:%M:%S%z) corresponding to each 10-second window, with shape (Nwindows ). • ship info: A group in the HDF5 file containing AIS metadata of the vessel used to generate the y label: – Vessel type, as a string value. – Vessel length, in meters. – Vessel beam, in meters. Note on Noisy Sensors Three sensors (indices 59, 60, and 61 in the X array) consistently exhibited high noise levels and have been forced to zero in the raw data. We recommend excluding these channels from feature matrices prior to training classification and regression models.
10
Table 2: Summary of released dataset contents. Component
Description
Dimensions
X y datetimes
3D array of energy-band features per channel Closest distance from the cable to any vessel [m] UTC timestamps (formatted as %Y-%m-%d %H:%M:%S%z) for each 10-s window AIS metadata for the vessel used to generate y labels Vessel type Vessel length [m] Vessel beam [m]
74,771 × 250 × 100 74,771 74,771
ship info
74,771 74,771 74,771
The public deposit does not include the original raw DAS recordings, complete AIS records, MMSI identifiers, or the precise geographical route of the cable because of data-owner, confidentiality and critical-infrastructure restrictions.
4
Technical Validation
The technical validation was designed to assess complementary aspects of the released dataset and its recommended evaluation protocol. We first define the baseline machine-learning framework and the ten-fold day-wise partitioning used throughout the validation. We then examine maritime-traffic diversity and environmental variability across these folds to assess whether they represent distinct physical vessel passages and heterogeneous acquisition conditions. Finally, we report baseline classification and regression results to verify that the released feature representation and AIS-derived labels retain information relevant to the two defined machine-learning tasks. These analyses are intended to support dataset quality and reuse rather than to provide an extensive comparison of modeling approaches, which is reported separately in [22].
4.1
Machine Learning Framework and Day-Wise Evaluation Protocol
We carried out the initial technical validation by applying an XGBoost ML algorithm to the two ML tasks defined in Section 2.5. The objective functions in the baseline XGBoost models are binary cross-entropy for classification and mean squared error for regression. We used gradient boosted trees, a learning rate η = 0.05, and a maximum tree depth of 10, with 500 boosting rounds. Our experimental approach is based on a rigorous k-fold cross-validation strategy, in which the dataset is divided into k equal parts (folds). A model is trained on k − 1 folds and tested on the remaining one, repeating this process k times so that each fold serves once as test data. The final performance metric is the average of the k test results, giving a more reliable estimate of the model’s performance, while keeping a strict separation between training and testing data. In our case, each fold corresponds to 1 full day of data, so that k = 10 and the cross-validation is day-wise based. The available source code at the companion GitHub repository also promotes the use of the cross-validation strategy. The vessel detection task was set with a distance threshold of 1,000 m, and the vessel distance estimation task evaluated two cases: unrestricted distance estimation, and distance estimation for vessels closer than 1,000 m to the cable. We exploited spatial redundancy in the DAS channels by averaging the feature vectors from NC contiguous channels in the monitored fiber-optic segment, which is given as an input to the XGBoost algorithm.
11
(a)
(b)
Figure 7: Directional traffic structure and recurrence of vessel identities across the day-wise folds: (a) Distribution of the lateral crossing positions of the 1,033 reconstructed passage events, separated according to the two dominant travel directions. (b) Distribution of the 565 unique vessel identities according to the number of day-wise folds in which each MMSI was observed.
4.2
Maritime Traffic Diversity Across the Day-Wise Folds
To characterize the local maritime traffic and assess the extent to which the day-wise folds represent distinct physical vessel passages within the monitored area rather than repeated observations of the same events, we analyzed vessel activity within a cable-adjacent region defined by expanding the bounding box of the monitored cable segment by 1 km in all directions. This analysis comprised 4, 290 original AIS observations (without interpolation), representing 565 unique MMSIs and 1, 033 reconstructed physical passage events1 . The complete daily folds contained between 97 and 122 passages. Traffic was approximately balanced between two opposite travel directions, with 507 and 526 passages, respectively. Passage directions were strongly bimodal: 95.4% of the reconstructed passages had AIS course-over-ground headings within two 30◦ -wide angular sectors centred on nearly opposite dominant travel directions. At a fixed reference transect approximately perpendicular to the dominant traffic direction, the corresponding directional traffic streams had median crossing positions separated by approximately 390 m, consistent with direction-dependent traffic lanes whose lateral distributions partially overlap as shown in Fig. 7.a. Most vessel identities were transient during the observation campaign, as shown by their distribution across the day-wise folds in Fig. 7.b. Of the 565 MMSIs, 322 occurred on one day and 198 on exactly two days; thus, 92.0% were present in no more than two daily folds. Among the 198 identities appearing on exactly two days, 192 generated only one passage on each day. For 189 of the 192 cases, the second passage occurred in the opposite traffic direction. These paired passages were separated by a median of 40.4 h and differed in lateral crossing position by a median of approximately 311 m (see distribution in Fig. 7.a). The identity recurrence observed between folds therefore primarily represents distinct outward and return passages rather than repeated measurements of equivalent traffic events. Only four of the 1,033 reconstructed passages crossed a daily fold boundary (over midnight), corresponding to 0.39% of all passage events and 16 of the 4,290 AIS observations. Among the test-fold MMSI occurrences for which the same MMSI also occurred in the training folds, 1.35% corresponded to the same continuous physical passage. After excluding these boundary events, the minimum temporal separation from a same-identity training passage had a median of 36.7 h, and 81.8% of the cases were separated by more than 24 h. These results indicate that the day-wise partitioning is effectively passage-disjoint and predominantly evaluates new physical traffic events under realistic maritime conditions. Although vessel identities naturally recur in an operational monitoring region, such recurrence rarely corresponds 1
The original MMSI identifiers were used only for this post hoc characterization of vessel recurrence and are not included in the released feature representation or provided as inputs to the machine-learning models.
12
to duplication of the same passage or adjacent DAS observations. The resulting folds therefore provide a robust and realistic basis for model validation, with negligible passage-level overlap and substantial variation in vessel identities, passage timing, travel direction, and crossing geometry.
4.3
Environmental Variability Across the Day-Wise Folds
To complement the maritime-traffic characterization, we examined the environmental conditions covered by the ten day-wise folds. Historical marine and meteorological data were retrieved through the Stormglass API at several query locations covering the monitored region. The variables considered were wave height, wave period, wind speed, water temperature, precipitation, and current speed. Fig. 8 shows the daily ranges at the closest location to the monitored region. Variables identified by the sg source correspond to the Stormglass aggregate source, whereas those identified by noaa correspond to NOAA-source values returned through the same API. Under the applicable data-use conditions, the original time-resolved StormGlass API responses cannot be publicly redistributed. We therefore release the derived file meteorological daily summary.csv, which contains, for each UTC day, the minimum and maximum values of the six variables shown in Fig. 8. These summaries are intended to document the environmental conditions represented in each fold, rather than to provide a replacement for the original meteorological data feed.
Figure 8: Daily minimum-to-maximum ranges of six environmental variables during the ten-day acquisition campaign. The daily ranges show substantial variation during the acquisition campaign. Across the ten days, wave height ranged from 0.10 to 1.31 m, wave period from 1.95 to 5.78 s, wind speed from 0.26 to 9.46 m/s, water temperature from 16 to 21.21 C, precipitation from 0 to 0.54 mm/h, and current speed from 0 to 1.07 m/s. Both the positions and widths of the daily intervals varied across folds. For example, the largest wave-height and wind-speed values occurred on 19 June, whereas the highest water temperature and longest wave period occurred on 25 June. Precipitation was absent during most daily intervals but reached measurable values on a subset of days. Consistent with this descriptive characterization, we carried out a statistical evaluation to assess whether the six meteorological variables differed significantly across the ten day-wise folds. Kruskal–Wallis tests, with Holm correction across the six variables, indicated significant day-to13
day differences in wave height (H(9) = 113.99, pHolm = 6.79 × 10−20 , ε2 = 0.458), wave period (H(9) = 124.92, pHolm = 5.25 × 10−22 , ε2 = 0.506), wind speed (H(9) = 60.39, pHolm = 1.13 × 10−9 , ε2 = 0.224), water temperature (H(9) = 210.12, pHolm = 1.50 × 10−39 , ε2 = 0.878), precipitation (H(9) = 74.33, pHolm = 4.29 × 10−12 , ε2 = 0.285), and current speed (H(9) = 178.36, pHolm = 5.57 × 10−33 , ε2 = 0.740). A PERMANOVA performed on the standardized values of the six variables also detected significant multivariate differences in the combined environmental conditions among days (pseudo-F = 24.01, R2 = 0.485, p = 1.0 × 10−4 ; 9,999 permutations). Taken together, these results show that the day-wise folds were acquired under measurably different marine and meteorological conditions rather than under a nearly stationary environmental state. This analysis is intended to characterize the environmental diversity of the released dataset and the recommended validation folds. It should not be interpreted as demonstrating that model performance is independent of meteorological conditions, nor as establishing representativeness across the full seasonal range of the deployment region.
4.4
Baseline Validation Results
The evaluation metrics used depended on the ML task: • For the classification task (vessel detection): Accuracy, and global and class-wise F1 scores, allowing performance for both proximity classes to be examined in the presence of class imbalance, with higher values meaning better performance. In our evaluation results (see Table 3), class 0 and class 1 cases refer to vessels closer or further than the distance threshold, respectively. • For the regression task (vessel distance estimation): Global Mean Absolute Error (MAE), which is computed considering all the available vessels in the dataset, and the below-1,000 m MAE, computed only for samples whose vessel-to-cable distance is below 1,000 m, to characterize estimation error in the proximity range most relevant to cable monitoring. Lower values indicate smaller estimation errors. Table 3 shows the performance metrics corresponding to NC ∈ {10, 50, 250} channels. To summarize, the classifier achieved a global F1 -score of 89.43% when averaging all 250 channels, with class-wise F1 -scores of 83.10% for the vessel-nearby class and 92.47% for the no-vessel-nearby class. For distance estimation, the corresponding global MAE was 829.90 m, while the MAE evaluated for samples below 1, 000 m was 171.01 m. Performance improved consistently as the number of spatial channels used for averaging increased. Table 3: Performance metrics for the vessel detection and distance estimation tasks under different spatial contexts (# channels refers to the number of channels considered for averaging).
# channels 10 50 250
Vessel detection task (distance threshold 1000 m) Accuracy Global F1 Class 0 F1 Class 1 F1 87.38% 87.01% 78.54% 91.06% 89.23% 88.99% 82.11% 92.29% 89.58% 89.43% 83.10% 92.47%
Distance estimation task Global MAE Below-1000 m MAE 979.84 m 196.89 m 882.79 m 191.33 m 829.90 m 171.01 m
The baseline results show that models trained on the released feature representation can discriminate the defined vessel-proximity classes and estimate vessel-to-cable distance under the reported evaluation protocol. These results provide an application-based validation of the information retained in the processed dataset. Additional experimental analyses using this dataset are reported in [22].
14
5
Usage Notes
The Marlinks-NS DAS dataset is intended to serve as a benchmark resource for developing and validating ML methods for vessel detection and vessel-to-cable distance estimation using submarine DAS data in submarine cable protection applications. The dataset is structured to enable reproducible experimentation and flexible adaptation to various modeling approaches. The spectral feature matrices can be directly used as input to ML models for the binary vessel detection or continuous vessel-distance regression tasks. For custom experiments, users may re-aggregate energy bands, apply normalization, or augment the data using temporal or spatial transformations. The AIS-derived distance labels may also be thresholded to define categorical proximity classes if required, including multi-class classification tasks. Simple Python examples for basic interaction with the dataset are distributed with the archived Zenodo release in src.zip. These scripts demonstrate how to inspect the HDF5 structure, load the complete feature arrays or selected slices, verify consistency between loading strategies, and generate day-wise (k-fold) cross-validation training and test partitions. Further software resources are described in Section 7.
6
Data Availability
The Marlinks-NS DAS dataset and its accompanying documentation are openly available in the Zenodo repository [30] at https://doi.org/10.5281/zenodo.15611778. The versioned deposit contains the complete processed dataset in HDF5 format (dataset sensor range 1440 1690 0.h5); dataset documentation in Markdown and PDF formats (README.md and README.pdf); supporting citation, licensing, provenance, funding, creators, changelog, DOI, daily ranges of environmental variables, and frequency-band metadata in misc.zip; and simple Python examples for inspecting, loading, checking, and partitioning the HDF5 data in src.zip. The data and accompanying documentation are released under the Creative Commons Attribution 4.0 International license.
7
Code Availability
A fixed snapshot of the basic Python data-handling scripts associated with the released dataset version is archived in the Zenodo record [30] within the src.zip archive. The companion GitHub repository provides extended data-handling, partitioning, visualization, and reproducibility tools, and will continue to be actively maintained and updated with corrections, documentation, and new functionality. Because the GitHub repository will evolve after publication, users should record and report the release, tag, or commit used in their analyses. The archived scripts require Python 3.8 or later and have been tested under Python 3.10 and 3.11, relying exclusively on widely available open-source Python packages. Dependencies for the extended software are documented in the GitHub repository. The released source code is distributed under the GNU General Public License v3.0. Users are encouraged to adapt and extend the provided scripts for their research needs. Contributions to the repository, including improvements, additional examples, and derived analysis tools, are very welcome through standard GitHub pull requests.
References [1] Naval News, “Seabed Cable Damaged in Latest Baltic CUI Incident,” Online, accessed June 2026, 2024. [2] Y. Zoria, “Finland-Germany submarine cable damaged again in Baltic Sea in possible sabotage act,” Online, accessed June 2026, 2025.
15
[3] Z. J. Spica, J. Ajo-Franklin, G. C. Beroza, B. Biondi, F. Cheng, B. Gaite, B. Luo, E. Martin, J. Shen, C. Thurber, L. Viens, H. Wang, A. Wuestefeld, H. Xiao, and T. Zhu, “PubDAS: A PUBlic Distributed Acoustic Sensing Datasets Repository for Geosciences,” Seismological Research Letters, vol. 94, no. 2A, pp. 983–998, 01 2023. [Online]. Available: https://doi.org/10.1785/0220220279 [4] W. Huang, S. Chen, Y. Wu, R. Li, T. Li, Y. Huang, X. Cao, and Z. Li, “DAShip: A Large-Scale Annotated Dataset for Ship Detection Using Distributed Acoustic Sensing Technique,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 18, pp. 4093–4107, 2025. [Online]. Available: https://doi.org/10.1109/jstars.2024.3525082 [5] ——, “DAShip dataset,” 2025. [Online]. Available: https://www.alipan.com/s/sTdL3zSRiPo [6] F. Cheng, B. Chi, N. J. Lindsey, T. C. Dawe, and J. B. Ajo-Franklin, “Utilizing distributed acoustic sensing and ocean bottom fiber optic cables for submarine structural characterization,” Scientific reports, vol. 11, no. 1, p. 5613, 2021. [Online]. Available: https://doi.org/10.1038/s41598-021-84845-y [7] F. Cheng, “Dataset from the mBARI DAS Project,” Dataset, 2020. [Online]. Available: https://doi.org/10.17605/OSF.IO/CN8XB [8] A. Sladen, D. Rivet, J. P. Ampuero, L. De Barros, Y. Hello, G. Calbris, and P. Lamare, “Distributed sensing of earthquakes and ocean-solid Earth interactions on seafloor telecom cables,” Nature communications, vol. 10, no. 1, p. 5777, 2019. [Online]. Available: https://doi.org/10.1038/s41467-019-13793-z [9] A. Sladen, “Dataset for the MEUST-NUMerEnv/KM3NeT DAS experiment Feb. 2018,” Dataset, 2019. [Online]. Available: https://doi.org/10.17605/OSF.IO/X6AWB [10] I. Lior, A. Sladen, D. Rivet, J.-P. Ampuero, Y. Hello, C. Becerril, H. F. Martins, P. Lamare, C. Jestin, S. Tsagkli et al., “On the detection capabilities of underwater distributed acoustic sensing,” Journal of Geophysical Research: Solid Earth, vol. 126, no. 3, p. e2020JB020925, 2021. [Online]. Available: https://doi.org/10.1029/2020JB020925 [11] I. Lior, “The Underwater DAS Detection Dataset,” Dataset, 2019. [Online]. Available: https://doi.org/10.17605/OSF.IO/4BJPH [12] K. Taweesintananon, M. Landrø, J. R. Potter, S. E. Johansen, R. A. Rørstadbotnen, L. Bouffaut, H. J. Kriesell, J. K. Brenne, A. Haukanes, O. Schjelderup, and F. Storvik, “Distributed acoustic sensing of ocean-bottom seismo-acoustics and distant storms: A case study from Svalbard, Norway,” GEOPHYSICS, vol. 88, no. 3, p. B135–B150, 2023. [Online]. Available: https://doi.org/10.1190/geo2022-0435.1 [13] K. Taweesintananon and M. Landrø, “Replication data for DAS4Microseism - Svalbard distributed acoustic sensing (DAS) strain data for oceanographic study,” 2022. [Online]. Available: https://doi.org/10.18710/VPRD2H [14] Q. Shi, E. F. Williams, B. P. Lipovsky, M. A. Denolle, W. S. Wilcock, D. S. Kelley, and K. Schoedl, “Multiplexed distributed acoustic sensing offshore central Oregon,” Seismological Research Letters, vol. 96, no. 2A, pp. 784–800, 2025. [Online]. Available: https://doi.org/10.1785/0220240460 [15] B. Lipovsky, E. Williams, and O. O. Initiative, “RAPID: Multiplexed Distributed Acoustic Sensing (DAS) at the Ocean Observatory Initiative (OOI) Regional Cabled Array (RCA),” Dataset, 2024. [Online]. Available: https://doi.org/10.58046/4WEF-A282
16
[16] Z. Song, X. Zeng, S. Ni, B. Chi, T. Xu, Z. Wei, W. Jiang, S. Chen, and J. Xie, “Near real-time in situ monitoring of nearshore ocean currents using distributed acoustic sensing on submarine fiber-optic cable,” Earth and Space Science, vol. 11, no. 9, p. e2024EA003572, 2024. [Online]. Available: https://doi.org/10.1029/2024EA003572 [17] ——, “Dataset in “Near real-time in-situ monitoring of nearshore ocean currents using Distributed Acoustic Sensing on submarine fiber-optic cable” ,” 2024. [Online]. Available: https://doi.org/10.5281/zenodo.13133835 [18] GEBCO Bathymetric Compilation Group, “GEBCO 2024 Grid [data set],” 2024. [Online]. Available: https://doi.org/10.5285/1c44ce99-0a0d-5f4f-e063-7086abc0ea0f [19] D. Rivet, B. de Cacqueray, A. Sladen, A. Roques, and G. Calbris, “Preliminary assessment of ship detection and trajectory evaluation using distributed acoustic sensing on an optical fiber telecom cable,” The Journal of the Acoustical Society of America, vol. 149, no. 4, pp. 2615–2627, 2021. [Online]. Available: https://doi.org/10.1121/10.0004129 [20] L. Thiem, S. Wienecke, K. Taweesintananon, M. Vaupel, and M. Landrø, “Ship noise characterization for marine traffic monitoring using distributed acoustic sensing,” in 2023 IEEE International Workshop on Metrology for the Sea; Learning to Measure Sea Health Parameters. IEEE, 2023, pp. 334–339. [Online]. Available: https://doi.org/10.1109/ metrosea58055.2023.10317227 [21] A. Tomasov, P. Zaviska, P. Dejdar, O. Klicnik, T. Horvath, and P. Munster, “Comprehensive dataset for event classification using distributed acoustic sensing (DAS) systems,” Scientific Data, vol. 12, no. 1, p. 793, 2025. [Online]. Available: https: //doi.org/10.1038/s41597-025-05088-4 [22] E. E. Ramirez-Torres, J. Macias-Guarasa, D. Pizarro-Perez, J. Tejedor, S. E. PalazuelosCagigas, P. J. Vidal-Moreno, S. Martin-Lopez, M. Gonzalez-Herraez, and R. Vanthillo, “Vessel detection and localization using distributed acoustic sensing in submarine optical fiber cables (accepted for publication),” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2026. [Online]. Available: https://doi.org/10.1109/JSTARS.2026.3716768 [23] Alcatel Submarine Networks, “OptoDAS interrogator,” Online, accessed June 2026. [24] H. Gabai and A. Eyal, “On the sensitivity of distributed acoustic sensing,” Optics letters, vol. 41, no. 24, pp. 5648–5651, 2016. [Online]. Available: https://doi.org/10.1364/OL.41.005648 [25] W. Zou, S. Yang, X. Long, and J. Chen, “Optical pulse compression reflectometry: proposal and proof-of-concept experiment,” Optics Express, vol. 23, no. 1, p. 512, 2015. [Online]. Available: https://doi.org/10.1364/oe.23.000512 [26] O. H. Waagaard, E. Rønnekleiv, A. Haukanes, F. Stabo-Eeg, D. Thingbø, S. Forbord, S. E. Aasen, and J. K. Brenne, “Real-time phase-recording DAS in 171 km low-loss fiber,” in Optical Fiber Sensors Conference 2020 Special Edition, ser. OFS. Optica Publishing Group, 2021, p. T2A.3. [Online]. Available: https://doi.org/10.1364/ofs.2020.t2a.3 [27] T. Emmens, C. Amrit, A. Abdi, and M. Ghosh, “The promises and perils of Automatic Identification System data,” Expert Systems with Applications, vol. 178, p. 114975, 2021. [Online]. Available: https://doi.org/10.1016/j.eswa.2021.114975 [28] S. Guo, J. Mou, L. Chen, and P. Chen, “Improved kinematic interpolation for AIS trajectory reconstruction,” Ocean Engineering, vol. 234, p. 109256, 2021. [Online]. Available: https://doi.org/10.1016/j.oceaneng.2021.109256
17
[29] R. Holman, H. Glover, M. Wengrove, M. Ifju, D. Honegger, and M. Haller, “Geolocation of Distributed Acoustic Sampling Channels Using X-Band Radar and Optical Remote Sensing,” Remote Sensing, vol. 17, no. 18, p. 3142, 2025. [Online]. Available: https://doi.org/10.3390/rs17183142 [30] E. E. Ramirez-Torres, J. Macias-Guarasa, D. Pizarro-Perez, J. Tejedor, S. E. PalazuelosCagigas, P. Vidal-Moreno, M. R. Fernández-Ruiz, S. Martin-Lopez, M. Gonzalez-Herraez, and R. Vanthillo, “Marlinks-NS DAS: Dataset for Vessel Detection and Distance Estimation Using Distributed Acoustic Sensing in Submarine Cables,” 2026. [Online]. Available: https://doi.org/10.5281/zenodo.15611778
Author Contributions All authors defined the released dataset, conceived the study, and designed the experimental methodology machine-learning tasks. R.V. coordinated the acquisition of and access to the DAS data, acquisition metadata, and preprocessing software. M.R.F.-R., S.M.-L. and M.G.-H. contributed to DAS preprocessing and optical-signal interpretation. E.E.R.-T., J.M.-G., D.P.-P. and S.E.P.-C. processed the AIS and geospatial information, synchronized it with the DAS measurements, and generated the vessel-distance labels. E.E.R.-T., D.P.-P. and J.M.-G. performed feature extraction, dataset curation, and design of the released HDF5 structure. E.E.R.-T., D.P.-P., J.M.-G., J.T. and P.J.V.M. developed the machine-learning software and carried out the technical validation. E.E.R.-T., J.M.-G. and S.E.P.-C. prepared and published the data and code repositories and, together, drafted the manuscript. J.M.-G., M.R.F.-R., S.M.-L. and M.G.-H. provided scientific supervision, project coordination, and funding acquisition. All authors reviewed and approved the manuscript.
Competing Interests The authors declare no competing interests.
Acknowledgments We gratefully acknowledge the computer resources at Artemisa, funded by the “European Union ERDF” and “Comunitat Valenciana” as well as the technical support provided by the “Instituto de Fı́sica Corpuscular”, IFIC (CSIC-UV). We also thank the cable monitoring operator and the cable owner for allowing data access under confidentiality requirements.
Funding This work has been partially supported by the “Spanish Ministry of Science and Innovation” MICIU/AEI/10.13039/501100011033, FEDER UE, and by the “European Union NextGeneration EU/PRTR” program under grants PSI (PLEC2021-007875), NeurEYE-UAH (PID2024-156576OBC31), SEASNAKE+ (PCI2023-145978-2, of the CETPartnership 2022 joint call), and MOTION (PID2022-140963OA-I00); by the “European Innovation Council” under grant ECSTATIC (101189595); and by the European Research Council under grant SENSE (101218803). The work of M.R.FR. was also supported by MCIN/AEI/10.13039/501100011033 and European Union “NextGenerationEU/PRTR” under grant RYC2021-032167-I. Open Access License. This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to
18
the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
19