OpenSAR Insight

OpenSAR Insight is an early-phase project developing ML-ready datasets and algorithms for direct insight generation from raw SAR data. This is a first step toward efficient onboard implementation of end-to-end SAR inference pipelines for low-latency applications.

The project is funded and supported by ESA's Φ-lab.

Datasets

We compiled novel datasets for three use cases: vessel detection, flood detection, and radio frequency interference (RFI) detection. All are built from Copernicus Sentinel-1 (S1A/S1B) acquisitions in Interferometric Wide (IW) mode, using both VV and VH polarisations.

Each dataset provides co-registered patches at multiple processing levels — Level-0 RAW, Single Look Complex (SLC), and Ground Range Detected (GRD) — with Level-0 extracted from the Level-1 products used. Labels are supplied as XML alongside the imagery, and are partially derived from SARFish/xView3, Kuro Siwo, and Aresys respectively.

L0, SLC and GRD counts are double the unique patch count because each patch is provided in both polarisations. Flood and RFI datasets include segmentation masks.

Vessel and flood patches are 512 × 512 px. RFI patches range from roughly 1400 × 1400 to 1600 × 1600 px, as they follow SLC burst boundaries. Train, validation and test splits are provided for every use case.

Coverage spans the Gulf of Guinea, Iceland, Denmark, Norway and Italy for vessel detection; Albania, Madagascar, India, Nicaragua, Pakistan, the Philippines, the United States, Lithuania, Sri Lanka and Peru for flood detection; and sites across Europe, the Middle East, East Asia, Africa and the Americas for RFI.

Models

For the vessel detection and RFI detection use cases, we developed two families of ML models: a large, unrestricted model for ground-based use, and a small, optimised model designed for onboard deployment. The small vessel detection model is knowledge-distilled from the large one.

• vessel-detection-model — YOLO-based
• rfi-detection-model — UNet-based

Codebase

The full codebase is at https://github.com/ESA-PhiLab/OpenSARInsight

It covers dataset generation from Sentinel-1 products, SAR preprocessing, model training and validation, and inference, with a Docker environment for reproducibility.

The code is released under the MIT License, except for components built on Ultralytics YOLO, which are distributed under AGPL-3.0. Licences for the datasets are given in their individual repositories.

Consortium

Indra Space leads the project, supporting the generation of open-source datasets and leading the development of ML models for extracting insights from raw SAR data. The consortium also includes the University of Alcalá (AES3 research group), INTA — the National Institute of Aerospace Technology (Spaceborne SAR Systems and Calibration Group), and ESA's Φ-lab.

• Indra Space
• University of Alcalá
• INTA
• ESA Φ-lab

Resources

• Project website
• Publications