PINGEcosystem/sss-crab-pot-detection-ds

🦀 Ghost Pot Side‑Scan Sonar Detection Dataset

Side‑Scan Sonar Imagery & Annotations for Derelict Crab Pot Detection

This dataset contains manually annotated side‑scan sonar (SSS) imagery collected across Delaware’s Inland Bays and Delaware Bay to support research on automated detection of derelict crab pots (“ghost pots”). It is designed for training and evaluating object‑detection models in turbid, shallow‑water environments where visual surveys are limited and acoustic mapping is essential.

The dataset accompanies the GhostVision project, an open‑source pipeline for near–real‑time detection and mapping of derelict fishing gear.

🤖 Related Models

The corresponding trained models are available at PINGEcosystem/gv-rf-detr, PINGEcosystem/gv-yolo12, and PINGEcosystem/gv-yolo26.

📜 Publication

In Progress

📦 Dataset Overview

• Total images: 6,674

• Sensor type: Consumer‑grade Humminbird side‑scan sonar

• Spatial coverage:

  • northern Rehoboth Bay near Dewey Beach
  • western Indian River Bay
  • southern Indian River Bay near White Creek

• Annotation format: JSON Lines (JSONL)

• Annotation type: Axis‑aligned bounding boxes

• Classes:

  • Crab-Pot — high confidence
  • Maybe-Crab-Pot — ambiguous or low‑confidence pot‑like target

• Coordinate system: Pixel coordinates in image space

File structure:

sss-crab-pot-detection
├── train/
│   ├── <image>.jpg
│   └── metadata.jsonl
├── valid/
│   ├── <image>.jpg
│   └── metadata.jsonl
├── train/
│   ├── <image>.jpg
│   └── metadata.jsonl

All annotations were created manually by trained analysts using high‑resolution sonar mosaics and cross‑validated with field notes and independent review.

🎯 Motivation

Derelict crab pots are a major source of ghost‑fishing mortality and benthic habitat damage. Existing removal programs rely on opportunistic sonar use during retrieval operations, meaning no formal mapping or archiving of pot locations exists. As a result, managers lack:

• basin‑wide estimates of derelict pot abundance
• spatial patterns of accumulation
• quantitative evaluation of removal efforts

This dataset enables reproducible, scalable research on automated detection, mapping, and monitoring of derelict fishing gear.

🗂️ Data Structure

This dataset uses JSON Lines (JSONL) format, where each line corresponds to a single image and all annotations associated with that image. This structure is natively supported by the Hugging Face Datasets library and enables efficient streaming and inspection.

🧾 JSONL Schema

Each row has the following structure:

{
  "file_name": "<image filename>",
  "objects": {
    "bbox": [ [x, y, width, height], ... ],
    "category": [ "<class>", ... ],
    "area": [ <area>, ... ]
  }
}

🏷️ Field Definitions

• file_name — relative path to the image file
• objects.bbox — list of bounding boxes in [x, y, width, height] format
• objects.category — list of class labels ("Crab-Pot" or "Maybe-Crab-Pot")
• objects.area — list of pixel areas for each bounding box

📭 Example: Image With No Detections

{
  "file_name": "Contact_334_sslo_png_jpg.rf.9c8a769f3c36deba4278527e714e09b7.jpg",
  "objects": {
    "bbox": [],
    "category": [],
    "area": []
  }
}

🖼️ Example: Image With Annotations

{
  "file_name": "Contact_203_sslo_png_jpg.rf.d9b44748a1bf7d9a30126e649741d745.jpg",
  "objects": {
    "bbox": [
      [309, 307, 26, 21],
      [196, 198, 29.5, 19]
    ],
    "category": ["Crab-Pot", "Crab-Pot"],
    "area": [546, 560.5]
  }
}

🏷️ Class Labels

• Crab-Pot — high-confidence derelict pot
• Maybe-Crab-Pot — ambiguous or low‑confidence pot‑like target

These are stored as strings for readability and compatibility with the Hugging Face viewer.

🚀 Loading the Dataset

from datasets import load_dataset

ds = load_dataset(PINGEcosystem/sss-crab-pot-detection)

# Get test dataset
ds = ds['test']

# Access bounding boxes for the first image
ds[0]["objects"]["bbox"]

🧪 Recommended Uses

This dataset is suitable for:

• Object detection (YOLO, DETR, RT‑DETR, RF‑DETR, etc.)
• Acoustic target classification
• Weakly supervised or semi‑supervised learning
• Spatiotemporal persistence modeling
• Benchmarking confidence/persistence fusion methods
• Research on ghost‑gear mapping and marine debris monitoring

⚠️ Limitations

• Sonar appearance varies with substrate, tow speed, and turbidity.
• Ambiguous targets are labeled as Maybe-Crab-Pot to avoid forcing false certainty.
• Not all pots are visible in every pass; some are buried or obscured.
• Dataset is region‑specific (Delaware), though methods generalize broadly.
• Users should consider domain adaptation when applying models to other regions or sonar systems.

🔗 Related Resources

• GhostVision project: PINGEcosystem/GhostVision
• RF-DETR model: PINGEcosystem/gv-rf-detr
• YOLO12 model: PINGEcosystem/gv-yolo12
• YOLO26 model: PINGEcosystem/gv-yolo26

📜 License

This dataset is released under the GPL license.

🙌 Acknowledgments

This dataset was developed with support from:

• University of Delaware -- Center for Coastal Sediments Hydrodynamics and Engineering Lab (CSHEL)
• Delaware Sea Grant
• 2024 Autonomous Systems Bootcamp
• NOAA's Project ABLE
• NOAA Marine Debris Program
• Delaware Department of Natural Resources and Environmental Control (DNREC)
• Community volunteers participating in ghost‑gear surveys