Portfolio demonstration of rotated-box detection on remote sensing data
Baseline for cross-domain evaluation (high-res RSDD-SAR → Sentinel-1 GRD)
Out-of-scope use
Do not use to accuse individual vessels of illegal activity
Do not use as sole evidence in any enforcement or legal decision
Outputs are candidate detections, not verified vessel identifications
Operational maritime surveillance requires multi-source fusion and human review
Training data
Dataset: RSDD-SAR (Radar Satellite Dataset for Ship Detection)
~7,000 SAR chip images (512 × 512 px, ~3 m ground resolution)
10,263 annotated ship instances with rotated bounding boxes
Annotations in long-edge convention: cx, cy, h, w, angle (radians, h ≥ w)
Source imagery from Gaofen-3 (C-band, HH/HV polarization)
Split used:
Train: 5,000 images (from ImageSets/train.txt)
Val / Test: 2,000 images (from ImageSets/test.txt — no separate val split in dataset; test used for both)
Test subsets: inshore (159 images), offshore (1,841 images)
Annotations were converted from VOC rotated-box XML to YOLOv11-OBB polygon format using
src/data/rsdd_to_yolo.py.
Long-edge enforcement was applied to all boxes before conversion.
Training procedure
Hyperparameter
Value
Base model
yolo11n-obb.pt (pretrained)
Epochs
100 (early-stopped at 75, patience 20)
Image size
512 × 512
Batch size
16
Optimizer
AdamW (Ultralytics default)
Hardware
Colab Pro T4 GPU
Training time
~2 hours
Evaluation results
Evaluated with Ultralytics .val() at IoU threshold 0.5. mAP@0.5:0.95 is the COCO-style
area under the IoU-threshold curve from 0.5 to 0.95 in 0.05 steps.
Split
mAP@0.5
mAP@0.5:0.95
Precision
Recall
N images
Overall test
0.938
0.640
0.933
0.871
2,000
Inshore
0.763
0.483
0.760
0.696
159
Offshore
0.971
0.673
0.954
0.931
1,841
Key finding: 21-point inshore/offshore gap (mAP@0.5: 0.763 vs 0.971).
The model handles isolated ships in open water well but struggles in port and
coastal environments due to clutter, infrastructure returns, and closely spaced vessels.
Limitations
Inshore/offshore gap: ~21-point mAP gap. Port clutter and coastal infrastructure
produce false positives and suppress true detections near shorelines.
Resolution mismatch: Trained on ~3 m RSDD-SAR; will degrade on Sentinel-1 GRD (~10 m).
Cross-domain evaluation is planned for Week 3 of the project.
Single class: No vessel-type classification (tanker, cargo, fishing, etc.).
Bright-object false positives: Oil platforms, wind turbines, breakwaters, and other
high-backscatter structures can produce spurious detections.
Polarization: Training data is primarily HH-polarized Gaofen-3. Performance on
other polarizations (VV, dual-pol) is untested.
Ethical considerations
Maritime surveillance is inherently dual-use. This model was built for research; outputs
should never be used without human review:
Detections are AIS-dark vessel candidates, not confirmed illegal ships
A vessel being absent from AIS does not mean it is acting illegally — AIS failures,
fishing vessels under tonnage thresholds, and legitimate AIS-off transit all produce
the same signature
False positive rates at Sentinel-1 resolution (~10 m) are non-trivial and have not
been characterized for the cross-domain case
Any operational use must involve multi-source fusion, domain experts, and legal oversight
How to use
python
1from huggingface_hub import hf_hub_download
2from ultralytics import YOLO
34# Downloads weights (~5.5 MB) and caches locally5weights = hf_hub_download(6 repo_id="tejassnaikk/rsdd-yolo11n-obb-v1",7 filename="best.pt",8)9model = YOLO(weights)1011# Run on a SAR image (grayscale or 3-channel, 512×512 recommended)12results = model("path/to/sar_chip.jpg", imgsz=512, conf=0.25)13results[0].show()# display with rotated boxes
Output format: results[0].obb contains oriented bounding boxes as xywhr tensors
(center x, center y, width, height, rotation in radians).
Citation / Project
@misc{sar-dark-ship-detection,
author = {Tejas Naik},
title = {SAR Dark Ship Detection},
year = {2025},
url = {https://github.com/tejassnaikk/sar-dark-ship-detection},
}
See the project repo for the
full pipeline including annotation conversion, evaluation harness, and Sentinel-1
cross-domain work.
