gavingavinchan/OpenDarkSTS-annotations

OpenDarkSTS Sentinel-2 Ship and Ship-to-Ship Annotation Dataset

Overview

This dataset contains 474 Sentinel-2 true-colour image chips of coastal and offshore waters, together with 2,636 human-reviewed bounding boxes. Each box marks either a single vessel, a pair of vessels lying alongside one another in the configuration of a ship-to-ship (STS) transfer, or a region that an earlier automatic detector flagged but that the reviewer judged to contain no vessel. The dataset is the annotation set described in the section "Annotation corpus" of the accompanying SNAME 2026 paper, and it was produced with the OpenDarkSTS review tool.

The images and labels are provided in the YOLO object-detection format, which can be used directly by Ultralytics YOLO training software, and with a per-image metadata file that allows the dataset to be loaded through the Hugging Face datasets library.

Related resources. Software: github.com/gavingavinchan/OpenDarkSTS. Default detector: gavingavinchan/OpenDarkSTS-YOLO11s. Companion dataset: gavingavinchan/OpenDarkSTS-v4-balanced.

Using the dataset

With YOLO training software. The file data.yaml defines the training, validation and test partitions and the two detector classes. Training may be started from the dataset folder, for example:

yolo detect train data=data.yaml model=yolo11s.pt imgsz=960

With the Hugging Face datasets library.

from datasets import load_dataset
ds = load_dataset("gavingavinchan/OpenDarkSTS-annotations")

Each record contains the image, its pixel dimensions, the region and acquisition date, an objects field holding the positive boxes, and a hard_negatives field holding the rejected regions. Boxes in the metadata files are expressed in pixels as [x_min, y_min, width, height]; boxes in the YOLO label files are expressed as normalised centre coordinates and dimensions.

Images

The images are rendered from Sentinel-2 MultiSpectral Instrument Level-2A products acquired between 2020 and 2025 and obtained from the Copernicus Data Space Ecosystem through Sentinel Hub. Each image is an 8-bit RGB true-colour rendering (bands B4, B3 and B2) at a ground sampling distance of approximately 10 m per pixel, stored as PNG. Image dimensions vary between 372 and 1,975 pixels per side, because chips were cut from scenes of different extents. Georeferencing, the near-infrared band and the exact acquisition time are not included.

The images cover four maritime areas:

Region	Images	ship_non_sts	ships_sts	neither
Gulf of Mexico, off the coast of Texas	186	1,072	36	51
Gulf of Suez	105	287	28	2
Singapore Strait	95	347	160	23
Strait of Gibraltar	88	409	180	41
Total	474	2,115	404	117

Annotation procedure

Each image was first processed by an earlier detector, which proposed candidate boxes. A reviewer then examined every image in the OpenDarkSTS manual review interface and accepted, resized, relabelled or deleted each proposal, and drew additional boxes where vessels had been missed. Of the 2,636 reviewed boxes, 2,047 (78%) originated as detector proposals that were subsequently corrected by the reviewer, and 589 (22%) were drawn manually. The origin of every box is recorded in the metadata. All images were reviewed by a single annotator.

Class definitions

Review label	Detector class (YOLO id)	Definition
ship_non_sts	ship (0)	A single vessel.
ships_sts	ships (1)	Two vessels lying alongside one another, in the configuration of a ship-to-ship transfer. One box encloses both hulls.
neither	none (excluded)	A region proposed by an earlier detector that the reviewer judged to contain no vessel (a hard negative).

The ships class describes a visual configuration only. It has not been confirmed against vessel tracking records and does not, by itself, establish that a cargo transfer took place.

Hard negatives

The 117 neither boxes are not written to the YOLO label files, so standard YOLO training treats them as background. They are retained in two places so that they can be used to train later detectors against the same false-positive patterns:

• the hard_negatives field of images/<split>/metadata.jsonl; and
• annotations/reviewed_boxes.csv, which lists all 2,636 reviewed boxes with their review label, normalised YOLO coordinates, origin, region and acquisition date. Rows with review_label = neither have an empty class_id.

One image contains no boxes, and two images contain only neither boxes; their YOLO label files are therefore empty.

Partitions and counts

The partition was generated by the OpenDarkSTS dataset export with its default scene-grouped rule: images sharing the same region and acquisition date are kept in the same partition, and groups are assigned to training, validation and test in the proportion 80:10:10 with random seed 42. This partition is provided for convenience. It is not the partition used to train the detectors reported in the paper (see the companion dataset opendarksts-v4-balanced).

Partition	Images	ship	ships	neither (hard negative)
Train	378	1,694	324	96
Validation	48	210	35	15
Test	48	211	45	6
Total	474	2,115	404	117

File layout

data.yaml                         YOLO dataset definition (classes 0: ship, 1: ships)
images/{train,val,test}/*.png     image chips
images/{train,val,test}/metadata.jsonl   per-image metadata, positive boxes and hard negatives
labels/{train,val,test}/*.txt     YOLO label files (positive classes only)
annotations/reviewed_boxes.csv    all reviewed boxes, including hard negatives

File names follow the source tile name where one was retained (for example 2020-01-24_scene1_tile_3_3.png) and otherwise a content-derived identifier (for example img_003386e7809374a3.png).

Source and attribution

Contains modified Copernicus Sentinel data 2020, 2021, 2024 and 2025, processed by the authors. Access to Sentinel data is governed by the Copernicus terms and conditions (https://dataspace.copernicus.eu/terms-and-conditions) and the Sentinel data legal notice. The bounding-box annotations were produced by the authors.

Licence

Creative Commons Attribution 4.0 International (CC BY 4.0). Users must also attribute the Copernicus Sentinel data as stated above. The OpenDarkSTS software and default detector are licensed separately under AGPL-3.0-only.

Citation

@inproceedings{opendarksts2026,
  title     = {OpenDarkSTS: Open-Source Dark Ship-to-Ship Transfers Detector},
  author    = {Chan, Pou Hei and Dey, Madhulika and Ansari, Sohail and Koola, Paul M.},
  booktitle = {Proceedings of the SNAME Maritime Convention},
  year      = {2026},
  note      = {Citation details to be completed on publication}
}

Limitations

• Incomplete labelling of single vessels. After a sufficient number of ship boxes had been collected, the reviewer stopped adding new ship boxes and concentrated on finding STS pairs. Some single vessels visible in the images are therefore not labelled, and the dataset should not be used to measure single-vessel recall without accounting for this. The ratio of ship to ships boxes (approximately 5:1) is a lower bound on the true imbalance.
• Single annotator. All boxes were reviewed by one person; inter-annotator agreement has not been measured.
• Visual labels only. The ships class has not been cross-checked against AIS records.
• Rendered imagery. The images are 8-bit true-colour renderings rather than calibrated surface reflectance, and contain no georeferencing.
• Boundary boxes. Four boxes extend marginally beyond the image boundary; they are reproduced exactly as reviewed.
• Overlap with the companion dataset. 222 of these images, with identical labels, also appear in opendarksts-v4-balanced, where they are assigned to partitions by a different rule. The two datasets should not be combined, nor should a model trained on one be evaluated on the other's test partition, without first removing shared images.