Model Card for OceanSAR-1-TenGeoP

Model Details

Model Description

OceanSAR-1-TenGeoP is a linear probing head for classifying ocean geophysical phenomena, built on top of the OceanSAR-1 foundation model. It leverages the powerful features extracted by OceanSAR-1 to accurately identify 10 different geophysical phenomena in Synthetic Aperture Radar (SAR) imagery.

• Developed by: Thomas Kerdreux, Alexandre Tuel @ Galeio
• Deployed by: Antoine Audras @ Galeio
• Model type: Linear Classification 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 labeled geophysical phenomena

Uses

Direct Use

This model is designed for automated classification of geophysical phenomena in SAR imagery over ocean surfaces. It can be used for:

• Rapid identification of ocean features in SAR data
• Monitoring of maritime environments
• Automated analysis of large SAR datasets
• Ocean science and research applications

Performance Results

The model achieves state-of-the-art performance on TenGeoP classification, with performance varying by backbone architecture:

How to Use

import torch
from transformers import AutoModelForImageClassification

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

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

# Extract features and classify geophysical phenomena
with torch.no_grad():
    outputs = oceansar(dummy_image)
    predicted_class = torch.argmax(outputs.logits, dim=1).item()

Training Details

Training Data

• Dataset: Sentinel-1 Wave Mode (WV) SAR images with labeled geophysical phenomena
• Labels: 10 classes of ocean geophysical phenomena
• Size: Balanced dataset across all classes
• Preprocessing: Same as base OceanSAR-1 model

Evaluation

Metrics

TenGeoP classification performance is evaluated using accuracy (%), achieving:

• 75.5% accuracy with ResNet50 backbone
• 78.6% accuracy with ViT-S/16 backbone
• 82.1% accuracy with ViT-S/8 backbone
• 83.6% accuracy with ViT-B/8 backbone

Comparison to Other Backbones

The model outperforms existing approaches:

• CROMA (ViT-B/8): 65.4% accuracy
• MoCo (ResNet50): 60.9% accuracy
• DeCUR (ResNet50): 58.3% accuracy
• DOFA (ViT-B/16): 58.4% accuracy
• DOFA (ViT-L/16): 63.4% accuracy
• SoftCon (ViT-S/14): 73.2% accuracy
• SoftCon (ViT-B/14): 74.8% accuracy

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.