opensar-insight/vessel-detection-dataset

Vessel Detection Dataset

Synthetic Aperture Radar (SAR) dataset for vessel detection, including dark/non-cooperative vessels not broadcasting AIS, developed as part of the OpenSAR Insight project. See the organization card for full project background, funding, and consortium details.

• —Codebase: https://github.com/ESA-PhiLab/OpenSARInsight
• —Project page: https://opensarinsightweb.web.uah.es/dark_vessel.html
• —Companion model: opensar-insight/vessel-detection-model

Overview

This dataset uses SAR images from the ESA Copernicus Sentinel-1 mission (Sentinel-1A and Sentinel-1B), which operate in all weather conditions, day and night. The selected Level 1 SAR products are Ground Range Detected (GRD) in wide interferometric mode (IW) and Single Look Complex (SLC) burst products.

Both available polarisation channels are used. VH (vertical-horizontal) polarisation is generally more conducive to ship detection, since it provides greater contrast between vessels and marine clutter, while VV (vertical-vertical) polarisation carries more information about sea-surface characteristics. A distinguishing feature of this dataset is the inclusion of Level 0 (L0, RAW) data extracted from the corresponding Level 1 products.

The dataset covers 15 SAR scenes. For each Level 1 product (GRD/SLC), the corresponding Level 2 Ocean (OCN) product was retrieved to add wind-condition information, complementing AIS ground truth from the xView3/SARFish reference database.

All L1 products were partitioned into non-overlapping 512x512 pixel blocks. Corresponding patches were extracted from GRD, SLC, L0 RAW, and L0 range-compressed data over the same geographic area, giving four co-registered patch types per block. Labels are provided in XML format.

Dataset organisation

The dataset is split into training, validation, and test folders. For each patch:

• —512x512 VV and VH L1 SLC patches under /patches
• —Corresponding VV and VH L1 GRD patches under /patches
• —Corresponding VV and VH L0 patches under /raw
• —Corresponding VV and VH L0 range-compressed patches under /range_compressed_rescaled
• —XML file with patch/scene information and vessel labels under /labels

Manual review for higher-quality subset

Some labels are imperfect — offset from the vessel centre, or too small to fully encompass it. A manual review was performed to filter for the highest-quality labels; this filtered subset is listed in yolo_dataset_filtered.csv and was used to train the models in opensar-insight/vessel-detection-model.

Downloading the data

pip install -U huggingface_hub

huggingface-cli login

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="opensar-insight/vessel-detection-dataset",
    repo_type="dataset",
    local_dir="./data"
)

License

MIT License. See the repository LICENSE for details on components of the broader codebase.

Citation

@misc{opensarinsight,
  title        = {OpenSAR Insight: ML-ready datasets and models for direct insight generation from raw SAR data},
  author       = {{Indra Space} and {INTA} and {Universidad de Alcal\'a de Henares}},
  howpublished = {\url{https://github.com/ESA-PhiLab/OpenSARInsight}},
  note         = {Funded by ESA \(\Phi\)-lab}
}