HF-EOLUS. Task 2. HF-Radar Wind Inversion Models and Results for VILA and PRIO Stations

This Zenodo record documents the methodology and results executed in pursuit of that goal. Readers seeking an extended methodological and results compendium can consult README.md, which delineates each pipeline, its configuration, and the evaluation metrics in detail.

This research combines high-frequency radar backscatter from the INTECMAR-maintained VILA and PRIO stations, Sentinel-1 Ocean (OCN) wind products, and in-situ measurements from the Puertos del Estado Vilano buoy. Together they recover 10 m wind vectors over the Galician continental shelf. The HF-radar streams arrive every 30 minutes, the buoy delivers hourly observations, and Sentinel-1 revisits supply intermittent orbital passes, furnishing a multi-scale temporal tapestry for model training. The initiative focuses on a reanalysis setting rather than real-time operations, enforcing the 5.7–17.8 m s⁻¹ validity gate imposed by the radar frequency throughout data preparation, model training, and evaluation. All signals are harmonised on a 20 km grid (10 km aggregation radius), with HF-radar echoes pivoted by Bragg peak, enriched with geometric descriptors, and tagged with maintenance-interval identifiers that support stability diagnostics and maintenance-aware normalisation.

A feed-forward artificial neural network delivers the inversion, sharing a backbone across three specialised heads: wind-speed regression, wind-direction regression (via sine/cosine outputs), and range classification. Hyper-parameter optimisation tunes depth, width, learning rates, and loss weights before definitive no-cross-validation training runs cement the baseline checkpoints. Downstream, fine-tuning explores three strategies—plain adaptation, L2-SP anchoring, and L2-SP augmented with knowledge distillation and rehearsal—to quantify how far knowledge can be transferred between the SAR grid and the buoy domain without eroding physical guardrails.

The study exposes three end-to-end pipelines. The SAR pipeline trains on Sentinel-1 supervision to capture spatial variability, evaluates both the home grid and the Vilano buoy, and measures the penalties incurred when the SAR backbone is adapted to buoy targets. The Vilano pipeline mirrors this flow, starting from the high-fidelity but point-scale buoy measurements before attempting to generalise across the grid. A third pipeline merges both corpora into a unified training/testing split to deliver a single cross-domain checkpoint. Complementary workflows extend the evaluation to a grid displaced 10 km east and north, stress-testing spatial robustness, and compute feature-importance diagnostics that highlight the dominant role of Bragg-peak power statistics, the stabilising influence of radial-velocity aggregates, and the conditional utility of geometric bearings when models venture outside their training domain.

Results confirm that each domain retains its native champion: the SAR baseline delivers 2.03 m s⁻¹ speed RMSE and 29.3° directional RMSE on the SAR test grid (in-range + class-match), the Vilano baseline reports 2.15 m s⁻¹ and 31.9° on the buoy test split, and the joint SAR+Vilano model balances both at 1.78 m s⁻¹ and 31.3° on the combined evaluation set. Cross-domain fine-tuning remains markedly asymmetric. Starting from the SAR checkpoint, plain adaptation trims the buoy matched RMSE from 2.84 to 2.30 m s⁻¹ (directional RMSE 28.6°) while the regularised continuations (L2-SP and L2-SP + KD) lower it further to 2.01 and 1.85 m s⁻¹ with ~30.6–30.8° angular errors; all three variants pay for these gains by inflating the SAR matched RMSE to 5.80, 5.56, and 3.66 m s⁻¹ respectively. In the opposite direction, Vilano→SAR continuations raise the grid matched RMSE above 2.9 m s⁻¹ and degrade buoy accuracy to 2.76–3.14 m s⁻¹ with >56° directional errors, leaving the native checkpoints as the most reliable options for buoy-centric deployments.

Reproducibility is encoded in hard-coded orchestration scripts (run_*_pipeline.sh) and mirrored directory structures archived in hf_eolus.zip, capturing configuration manifests, training checkpoints, fine-tuning artefacts, inference metrics, and stage-specific reports. Every materialised dataset passes through GeoParquet consolidation and STAC cataloguing, with the catalogues bundled under catalogs.zip, both preserving the original repository layout so that provenance, spatial metadata, and maintenance-aware normalisation parameters remain auditable. The combined workflow therefore constitutes a physically grounded, fully documented benchmark for HF-radar wind inversion on the Galician shelf, clarifying the trade-offs between spatial coverage and point accuracy when transferring skill across observation domains.

The software toolkit underpinning this research is documented in Herrera Cortijo, J. L., Fernández-Baladrón, A., Rosón, G., Gil Coto, M., Dubert, J., & Varela Benvenuto, R. (2025). HF-EOLUS HF-Radar Wind Inversion Toolkit for Artificial Neural Networks Training and Inference (v0.1.0). Zenodo. https://doi.org/10.5281/zenodo.17369264.

Inputs are documented in:

• Herrera Cortijo, J. L., Fernández-Baladrón, A., Rosón, G., Gil Coto, M., Dubert, J., & Varela Benvenuto, R. (2025). Project HF‑EOLUS. Task 1. Puertos del Estado Vilano Buoy Data Bundle (GeoParquet + STAC) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17098038
• Herrera Cortijo, J. L., Fernández-Baladrón, A., Rosón, G., Gil Coto, M., Dubert, J., & Varela Benvenuto, R. (2025). Project HF-EOLUS. Task 2. Aggregated SAR and HF-Radar Radial Metrics for Wind-Inversion Model Training [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17115413.

What's New

A buoy wind height correction stage has been incorporated through the dockerised wrapper scripts/aggregation/apply_buoy_wind_height_correction.sh, which applies the neutral logarithmic profile to translate the Vilano buoy winds from their native 3 m measurement height to the 10 m reference used across the SAR products. All affected artificial neural network checkpoints were retrained with the corrected buoy series, and the resulting metrics remain within the previously reported uncertainty envelope—no substantial shifts were detected in RMSE, directional error, or range-classification accuracy.

Acknowledgements

This work has been funded by the HF-EOLUS project (TED2021-129551B-I00), financed by MICIU/AEI /10.13039/501100011033 and by the European Union NextGenerationEU/PRTR - BDNS 598843 - Component 17 - Investment I3. Members of the Marine Research Centre (CIM) of the University of Vigo have participated in the development of this repository.

We thank Puertos del Estado for making the Vilano buoy data publicly available through Portus (https://portus.puertos.es).

Aggregated HF-Radar data was derived from spectra from INTECMAR's VILA and PRIO HF-Radar stations, between 2011-09-30 and 2023-11-23 that have been transferred free of charge by the Observatorio Costeiro da Xunta de Galicia (https://www.observatoriocosteiro.gal) for their use. This Observatory is not responsible for the use of these data nor is it linked to the conclusions drawn with them. The Costeiro da Xunta de Galicia Observatory is part of the RAIA Observatory (http://www.marnaraia.org).

Disclaimer

This data and software is provided "as is", without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose, and noninfringement. In no event shall the authors or copyright holders be liable for any claim, damages, or other liability, whether in an action of contract, tort, or otherwise, arising from, out of, or in connection with the software or the use or other dealings in the software.