gavingavinchan/OpenDarkSTS-YOLO11s

Default model: SNAME 2026 YOLO11s

This detector identifies candidate pairs of ships in approximately 10 m/pixel, true-colour satellite imagery for human review. Classes: 0: ship, 1: ships. A visible pair does not establish a transfer, vessel identity, cargo, transponder state, or legality. opendarksts init downloads it; opendarksts init --weights PATH_TO_THIS_CHECKPOINT installs a local copy.

Checkpoint: opendarksts_yolo11s_v4_best.pt SHA-256: cf11d54de2aa2ba3fa74e23d1118482a281b62cd29a95b2b1e58a06e92b261ea Download (Hugging Face): https://huggingface.co/gavingavinchan/OpenDarkSTS-YOLO11s/resolve/main/opendarksts_yolo11s_v4_best.pt

The checkpoint is the YOLO11s headline detector presented in the SNAME 2026 paper, trained on madhulika_rectbbox_plus_manual_sts_balanced_v4. This balanced dataset contains 1,991 instances of each class, combining the pre-augmented Madhulika rectbbox source set with manually reviewed Sentinel-2 images and train-only manual STS-crop augmentation. Source metadata identifies Roboflow project fleetdatascience/data-science-dark-fleet, version 7, as Public Domain; the source-set author has agreed to its publication.

Held-out test results at image size 960, reproduced exactly from the paper's "Headline detector --- YOLO11s per-class results" section and cross-architecture table (21 test images):

Class	Instances	Precision	Recall	mAP50	mAP50–95
ship	62	0.812	0.806	0.869	0.512
ships	30	0.849	0.833	0.920	0.575
all	92	0.830	0.820	0.895	0.544

YOLO11s ranked highest by overall mAP50 among the eleven compared architectures; YOLOv12n had the highest mAP50–95 (0.572). These are recorded test metrics, not a fresh independent benchmark. Operational inference uses the checkpoint's training image size (960) and a default confidence of 0.25; these metrics do not measure the accuracy of the whole operational pipeline or of the dark-status labels.

Limitations: the test split contains only 21 images. The v4 dataset derives from a pre-augmented source set, so some test images share source scenes with training images. These results therefore do not establish generalization to independent scenes. Small or occluded hulls, cloud, wakes, ports, coarse imagery and unfamiliar regions can produce misses or false positives. Overlapping boxes or tiles can produce repeated candidates. AIS comparison uses any valid position within the radius on the same UTC day, with no acquisition-time window, track association or hull identification. Missing or invalid AIS remains unknown. Coverage is US AIS.

Training: COCO-pretrained yolo11s.pt; 300 epochs, image size 960, batch 8, seed 42, patience 60, Ultralytics 8.4.23. See weight_recreation.md for the command, source checkpoint and reproducibility gaps. The clean-split YOLO26n checkpoint remains an alternative documented in docs/yolo_weight_recreation.md; its scores use a different dataset and must not be substituted for the paper's YOLO11s results. Licence: AGPL-3.0-only, as for the project and its Ultralytics dependency.

Use

This checkpoint is the default detector of OpenDarkSTS; opendarksts init downloads it and verifies the SHA-256. Training data: gavingavinchan/OpenDarkSTS-v4-balanced.

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}
}