Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OceanSAR-2: A Universal Feature Extractor for SAR Ocean Observation

Alexandre Tuel, Thomas Kerdreux, Quentin Febvre, Alexis Mouche, Antoine Grouazel, Jean-Renaud Miadana, Antoine Audras, Chen Wang, Bertrand Chapron

TL;DR

OceanSAR-2 addresses the need for scalable, task-agnostic representations in SAR-based ocean observation where labeled data are scarce. It leverages a $DINOv2$-based self-supervised framework with $\sigma^0$-normalized inputs, an $iBOT$ local-patch loss, and a $KoLeo$ regularizer, plus dynamic data curation. The study demonstrates strong zero-shot transfer across classification, regression (SWH, wind), and iceberg detection using a compact ~21M-parameter backbone and 384-d embeddings. It also introduces a living benchmark suite (TenGeoP, WV-SWH, WV-wind, YOLOIB) to standardize evaluation and accelerate ocean-SAR research.

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.

OceanSAR-2: A Universal Feature Extractor for SAR Ocean Observation

TL;DR

OceanSAR-2 addresses the need for scalable, task-agnostic representations in SAR-based ocean observation where labeled data are scarce. It leverages a -based self-supervised framework with -normalized inputs, an local-patch loss, and a regularizer, plus dynamic data curation. The study demonstrates strong zero-shot transfer across classification, regression (SWH, wind), and iceberg detection using a compact ~21M-parameter backbone and 384-d embeddings. It also introduces a living benchmark suite (TenGeoP, WV-SWH, WV-wind, YOLOIB) to standardize evaluation and accelerate ocean-SAR research.

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.
Paper Structure (15 sections, 1 figure, 3 tables)

This paper contains 15 sections, 1 figure, 3 tables.

Table of Contents

  1. Introduction
  2. Building OceanSAR-2
  3. Self-supervised training strategy
  4. Training data and dataset curation
  5. Architecture
  6. A benchmarking framework for SAR ocean observation
  7. Current state
  8. Model benchmarking vs. fine-tuning
  9. Selected benchmarks
  10. TenGeoP
  11. WV-SWH
  12. WV-wind
  13. YOLOIB
  14. Results
  15. Discussion and conclusion

Figures (1)

  • Figure 1: Example of Sentinel-1 WV images (left column; top: sea-ice, middle: rain cells; bottom: icebergs) with corresponding OceanSAR-2 feature similarity maps (right column). The reference patch to compute the similarity is indicated by a red dot in both panels.