Model Card for OceanSAR-1-Wind

Model Description

OceanSAR-1-Wind is a linear probing head for wind speed prediction built on top of the OceanSAR-1 foundation model. It leverages the powerful features extracted by OceanSAR-1 to accurately predict wind speed from Synthetic Aperture Radar (SAR) imagery.

• Developed by: Thomas Kerdreux, Alexandre Tuel @ Galeio
• Deployed by: Antoine Audras @ Galeio
• Model type: Linear Regression Head on Vision Foundation Model
• License: Apache License 2.0
• Base model: OceanSAR-1 (ResNet50/ViT variants)
• Training data: Sentinel-1 Wave Mode (WV) SAR images with collocated wind speed measurements

Uses

Direct Use

This model is designed for wind speed prediction from SAR imagery, particularly over ocean surfaces. It can be used for:

• Near-real-time wind speed estimation from SAR images
• Assimilation into meteorological models
• Marine weather forecasting
• Offshore operations planning

Performance Results

The model achieves state-of-the-art linear probing performances on wind speed prediction, with performance varying by backbone architecture:

Backbone	Wind RMSE (m/s)
ResNet50	1.62
ViT-S/16	1.39
ViT-S/8	1.38
ViT-B/8	1.37

How to Use

import torch
from transformers import AutoModelForImageClassification

# Load the foundation model and wind prediction head
oceansar = AutoModelForImageClassification.from_pretrained("galeio-research/OceanSAR-1-wind")

# Prepare your SAR image (should be single-channel VV polarization)
dummy_image = torch.randn(1, 1, 256, 256)  # (C, H, W)

# Extract features and predict wind speed
with torch.no_grad():
    wind_speed = oceansar(dummy_image).logits  # Output in m/s

Training Details

Training Data

• Dataset: Sentinel-1 Wave Mode (WV) SAR images with collocated wind speed measurements
• Source: Wind speed measurements from scatterometer and buoy data
• Preprocessing: Same as base OceanSAR-1 model

Evaluation

Metrics

Wind speed prediction performance is evaluated using Root Mean Square Error (RMSE), achieving:

• 1.62 m/s RMSE with ResNet50 backbone
• 1.39 m/s RMSE with ViT-S/16 backbone
• 1.38 m/s RMSE with ViT-S/8 backbone
• 1.37 m/s RMSE with ViT-B/8 backbone

Comparison to Other Backbones

The model outperforms existing approaches:

• MoCo: 1.80 m/s RMSE
• DeCUR: 1.93 m/s RMSE
• SoftCon ViT-S/14: 1.98 m/s RMSE
• SoftCon ViT-B/14: 1.95 m/s RMSE

Technical Specifications

Hardware Requirements

• Same as base model
• Minimal additional computational cost for inference

Dependencies

• PyTorch >= 1.8.0
• Transformers >= 4.30.0
• Base OceanSAR-1 model

Input Specifications

• Same as base OceanSAR-1 model
• Single channel (VV polarization) SAR images
• 256x256 pixel resolution

Citation

BibTeX:

@article{kerdreux2025efficientselfsupervisedlearningearth,
  title={Efficient Self-Supervised Learning for Earth Observation via Dynamic Dataset Curation},
  author={Kerdreux, Thomas and Tuel, Alexandre and Febvre, Quentin and Mouche, Alexis and Chapron, Bertrand},
  journal={arXiv preprint arXiv:2504.06962},
  year={2025},
  eprint={2504.06962},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2504.06962},
}

Acknowledgements

This work was granted access to the HPC resources of IDRIS and TGCC under the allocation 2025-[A0171015666] made by GENCI.