Dataset DOI: 10.5061/dryad.brv15dvnz

Description of the data and file structure

Files and variables

DAS and mooring data are provided in netCDF (.nc) format and contain self-describing global attributes, variable descriptions, and units.

oliktok_das_mooring_hourly_dataset.nc

This dataset contains hourly DAS features (depth, cosine-square-wave direction, strain-rate spectra) and target mooring seafloor pressure spectra for training the spectral neural network model. Hourly strain rate and pressure spectral density estimates were computed from demeaned DAS and mooring records, respectively. Spectra were estimated using Welch's method with 512-point (256-s) Hann-windowed segments and 50% overlap. Frequencies above 0.5 Hz were trimmed to remove much of the noise floor, and every 4 adjacent frequencies were averaged to increase the spectral degrees of freedom (reduce noise). The resulting spectra span 0.0078 Hz to 0.49 Hz in 32 frequency bins. All strain rate spectra within 1 km of each mooring site are included (61 channels per site). Depths from NOAA navigation maps described in Baker & Abbott (2022) were interpolated at each channel. Cosine-squared wave directions were estimated by fitting the ocean surface gravity wave dispersion relationship to strain rate frequency-wavenumber (f-k) spectra created from 1 km of channels centered on each site, and the results were broadcast to all 61 channels at each site.

oliktok_das_mooring_half-hourly_dataset.nc

This dataset contains half-hourly DAS strain rate f-k spectra and target mooring seafloor pressure spectra required to train the f-k convolutional neural network model. Strain rate f-k spectra were estimated using two-dimensional arrays of raw strain rate from 2 km of DAS channels collected over 30 min. Data were segmented along the time axis into 600-point (300-s) segments with 50% overlap and were tapered with a 2D Hann window function prior to computing 2D Fourier transforms and averaging. Frequencies above 0.5 Hz were trimmed, and every four frequencies and three wavenumbers were merged. Final strain rate f-k spectra were interpolated onto a 32x32 grid of frequencies from 0.01 to 0.50 Hz and wavenumbers from +/- 0.00096 to 0.03 m^(-1), where negative wavenumbers indicate offshore wave propagation. Mooring seafloor pressure spectra were estimated using the same methods as oliktok_das_mooring_hourly_dataset.nc but were interpolated onto the same frequency bins used in the f-k spectra.

oliktok_das_hourly_along_cable_dataset.nc

This dataset contains hourly DAS features (depth, cosine square wave direction, strain rate spectra) in 2-km along-cable segments with 50% overlap (for a total of 25 segments, or artificial "sites"). Data were processed using the same methods as oliktok_das_mooring_hourly_dataset.nc, except DAS strain rate spectra, depth, and cosine squared wave direction were averaged over all 123 channels within each segment.

oliktok_das_half-hourly_along_cable_dataset.nc

This dataset contains half-hourly DAS strain rate f-k spectra in 2-km along-cable segments with 50% overlap (for a total of 25 segments, or artificial "sites"). Data were processed using the same methods as oliktok_das_mooring_half-hourly_dataset.nc.

model_states.zip

This file contains the trained spectral neural network and f-k convolutional neural network model states and fitted min-max normalizations required to apply the models. States are provided in PyTorch state dict format (.pth). Example code for loading the states and reconstructing the model and normalizations are provided in a separate repository hosted on Zenodo.

The .zip folder includes the following files:

• fk_convolutional_neural_network_hyperparameters.pth: Hyperparameters required to reconstruct the f-k CNN model.
• fk_convolutional_neural_network_model.pth: Trained f-k CNN model state.
• fk_convolutional_neural_network_spectral_feature_norm.pth: Fitted f-k CNN feature (f-k spectra) min-max normalization.
• fk_convolutional_neural_network_target_norm.pt: Fitted f-k CNN target (seafloor pressure spectra) min-max normalization.
• spectral_neural_network_hyperparameters.pth: Hyperparameters required to reconstruct the spectral NN model.
• spectral_neural_network_model.pth: Trained spectral NN model state.
• spectral_neural_network_scalar_feature_norm.pth: Fitted spectral NN scalar feature (depth, cosine squared wave direction) min-max normalization.
• spectral_neural_network_spectral_feature_norm.pth: Fitted spectral NN spectral feature (strain rate spectra) min-max normalization.
• spectral_neural_network_target_norm.pth: Fitted spectral NN target (seafloor pressure spectra) min-max normalization.

References

Baker, M. G., & Abbott, R. E. (2022). Rapid Refreezing of a Marginal Ice Zone Across a Seafloor Distributed Acoustic Sensor. Geophysical Research Letters, 49(24). https://doi.org/10.1029/2022GL099880

Code/software

Example code for training and applying the models is organized into Python notebooks and can be accessed via GitHub at https://github.com/jacobrdavis/neural_network-based_methods_for_das_ocean_surface_wave_measurement or via the Zenodo archive at https://doi.org/10.5281/zenodo.17381091.

NetCDF files can be opened using any software that supports the netCDF file format (e.g., the Xarray and netCDF4 Python packages or MATLAB's ncread function).  See https://www.unidata.ucar.edu/software/netcdf/ for more information.

Model state files (.pth) can be loaded using PyTorch. PyTorch v2.9.0 state dict (.pth) files are serialized Python OrderedDicts and can be opened using torch.load.

Example notebooks and code to reconstruct the models are provided in a separate repository hosted on Zenodo.

Access information

Other publicly accessible locations of the data:

• The complete April to September 2023 mooring dataset is archived on the Arctic Data center at https://doi.org/10.18739/A24J0B01J (Thomson and Smith, 2024)

References:

Thomson, J., & Smith, M. (2024). Moorings along the seafloor cable route extending offshore from Oliktok Point, Alaska, from April to September of 2023 [Dataset]. Arctic Data Center. https://doi.org/doi:10.18739/A24J0B01J