[Paper Review] OceanSAR-2: A Universal Feature Extractor for SAR Ocean Observation

Alexandre Tuel, Thomas Kerdreux+7 more|Archimer (Ifremer)|Jan 12, 2026

Ocean Waves and Remote Sensing0 citations

TL;DR

OceanSAR-2 is a compact DINOv2-based foundation model trained on calibrated sigma0 Sentinel-1 Wave Mode data, acting as a universal SAR ocean feature extractor with strong zero-shot and fine-tuning transfer across classification, regression, and detection tasks; it includes a standardized ocean SAR benchmarking framework.

ABSTRACT

We present OceanSAR-2, the second generation of our foundation model for SAR-based ocean observation. Building on our earlier release, which pioneered self-supervised learning on Sentinel-1 Wave Mode data, OceanSAR-2 relies on improved SSL training and dynamic data curation strategies, which enhances performance while reducing training cost. OceanSAR-2 demonstrates strong transfer performance across downstream tasks, including geophysical pattern classification, ocean surface wind vector and significant wave height estimation, and iceberg detection. We release standardized benchmark datasets, providing a foundation for systematic evaluation and advancement of SAR models for ocean applications.

Motivation & Objective

• Motivate the need for universal, SAR-specific foundation models for ocean observation due to data scarcity and ocean-specific imaging physics.
• Develop OceanSAR-2, a compact, sigma0-native self-supervised model, improving transfer to downstream ocean tasks while reducing training cost.
• Propose standardized, evolving benchmarking datasets to enable fair cross-model comparison for SAR ocean applications.
• Demonstrate the model’s transfer capabilities across pattern classification, geophysical parameter estimation, and iceberg detection.

Proposed method

• Adopt DINOv2-based self-supervised pretraining to capture both global image structure and local patch information (iBOT loss) with KoLeo regularization for diverse representations.
• Use calibrated sigma0 backscatter as input instead of raw DN amplitudes to improve feature consistency across acquisitions.
• Implement dynamic data pruning to reduce redundancy and promote diversity of training samples.
• Maintain a Vision Transformer backbone (patch size 16x16) producing embeddings plus a class token for image-level representation.
• Train on Sentinel-1 Wave Mode SAR data with a relatively small model size (≈21M parameters) to achieve strong transfer while keeping costs low.
• Provide standardized downstream benchmarks spanning classification, regression, and object detection for ocean SAR.

Experimental results

Research questions

• RQ1 Can a sigma0-native, self-supervised SAR model learn transferable representations for diverse ocean phenomena without labeled data?
• RQ2 How does OceanSAR-2 perform in zero-shot and fine-tuning settings across classification, regression (SWH, wind), and iceberg detection tasks?
• RQ3 Do standardized, living benchmarks enable meaningful comparison and progress tracking for ocean-SAR foundation models?
• RQ4 What is the trade-off between model size, training cost, and downstream performance for ocean SAR foundation models?

Key findings

• In zero-shot mode, OceanSAR-2 achieves top performance on all benchmarks compared to TerraMind, WV-Net, and DINOv3 (TenGeoP 94.0% accuracy, SWH 0.52 m, Wspd 1.32 m/s).
• In MLFine-tuning mode, OceanSAR-2 delivers highest or near-highest results on most tasks, with TenGeoP 98.5% accuracy, SWH 0.40 m, Wspd 1.01 m/s, Wdir 16.9 degrees, and YOLOIB 0.865 F1.
• OceanSAR-2 requires far less parameters (21M) and training cost than large models like DINOv3 (300M) while maintaining competitive performance.
• WV-Net is highly competitive on SWH (0.427–0.64 m) and wind (1.23–1.71 m/s) due to input normalization strategies (SSR) that emphasize wave modulation dynamics.
• TerraMind underperforms across tasks, likely due to lack of ocean-SAR pretraining data and MAE-based design.
• The study demonstrates that a compact, sigma0-native backbone can act as a universal feature extractor for SAR ocean imagery across classification, regression, and detection.