A Foundation Model for SAR Ocean Observation

Event: CVPR 2025 · Duration: 9 min · ▶ Watch on YouTube

Abstract

Self-supervised learning is transforming Earth observation by enabling the training of powerful foundation models on vast amounts of unlabeled satellite imagery. However, Earth observation data, particularly Synthetic Aperture Radar (SAR) imagery, presents a unique challenge due to its extreme redundancy, making naive training inefficient and prone to biased models. This work introduces a novel dynamic dataset pruning strategy that intelligently refines the training dataset during the self-supervised learning process, eliminating the need for pre-existing feature extractors or extensive metadata. Applied to Sentinel-1 wave mode SAR imagery over oceans, this method leads to the creation of OceanSAR-1, a specialized foundation model that demonstrates improved computational efficiency and superior performance across various ocean-specific downstream tasks.

Speakers

  • T. Kerduell — Galileo, Paris, France / Ifremer, UMR CNRS Lops, Brest, France
  • A. Tuel — Galileo, Paris, France / Ifremer, UMR CNRS Lops, Brest, France
  • Q. Febvre — Galileo, Paris, France / Ifremer, UMR CNRS Lops, Brest, France
  • A. Mouche — Galileo, Paris, France / Ifremer, UMR CNRS Lops, Brest, France
  • B. Chapron — Galileo, Paris, France / Ifremer, UMR CNRS Lops, Brest, France

Talks (1)

  • 00:00:00 — T. Kerduell: A Foundation Model for SAR Ocean Observation
    • This talk introduces a dynamic dataset pruning strategy for self-supervised learning on highly redundant Earth observation data, specifically focusing on Sentinel-1 SAR imagery over oceans, to create a specialized foundation model called OceanSAR-1.

Key Takeaways

  • Dynamic dataset pruning effectively addresses data redundancy in Earth observation, boosting both computational efficiency and feature quality for self-supervised learning.
  • The proposed method, applied to Sentinel-1 SAR ocean imagery, results in OceanSAR-1, a specialized foundation model that consistently outperforms existing SAR foundation models on ocean-specific tasks.
  • Smart data curation strategies, such as dynamic pruning, enable focused unimodal models to achieve superior performance compared to larger, more general multimodal models, especially for specialized satellite modalities.
  • The approach offers a fully self-supervised training strategy, eliminating the need for pre-existing feature extractors or extensive metadata for dataset curation, making it highly adaptable to new domains and sensor types.

Methods / Models / Datasets Mentioned

  • DINO
  • OceanSAR-1
  • Sentinel-1
  • TenGeoP
  • ERA5
  • ImageNet

Topics

Self-supervised learning · Earth observation · SAR imagery · Ocean observation · Dynamic dataset pruning · Foundation models · Data redundancy · Computational efficiency · Feature quality

Notes

Open for commentary — connections to other work, critiques, follow-up reading.