Dinoman1221/sonarvision-multisource-v6

🌊 KADAL Multi-Source Side-Scan Sonar Dataset (v6)

SIH 2026 | SIH26057 (Ministry of Earth Sciences / NIOT) — AI-Powered Automated Underwater Marine Debris and Anomaly Detection System using Side-Scan Sonar Imagery.

Official Project Source Code & Documentation:
🔗 https://github.com/Dinoman67/sonarvision
(Refer to the GitHub repository for preprocessing scripts, augmentations, model weights, edge dashboard, and optional supporting features such as C2 (NMEA/KML) exports, XTF/slant-range helpers, and a simulated live-waterfall demo view)

---

📌 Dataset Summary

The KADAL Multi-Source v6 Dataset is an acoustic object detection benchmark specifically curated for Side-Scan Sonar (SSS) imagery. It combines high-frequency acoustic surveys from multiple maritime sources into a unified, leak-free YOLO annotation format with 4 tactical object classes:

1. unknown_debris: Marine debris, discarded industrial items, containers, scrap, and navigation hazards — including net-like entangled debris (no open ghost-net benchmark exists; cf. GhostNetZero's private 412-segment set, so nets are covered here rather than as a separate class).
2. airplane: Submerged aircraft fuselages, wings, and wreckage.
3. mine: Underwater naval bottom mines, moored/tethered mine shapes, and acoustic mine-like objects (NOMBOs).
4. wreck: Sunken vessel hulls, structural wrecks, and maritime heritage sites.

---

📊 Dataset Statistics & Splits

The dataset comprises 5,558 high-resolution side-scan sonar frames organized into strict train, validation, and test splits with zero file overlap across splits (debris train/test share survey passes — frame-level split, not pass-level; see Limitations in the GitHub repo reports/debris_feature_learning_report.md):

Split	Images	Background / Negative Frames	Description
Train	4,033	~2,550	Comprehensive multi-condition training partition
Validation	563	381	Hyperparameter tuning and model checkpoint selection
Test (Held-out)	962	805	Completely unseen test track holdout for unbiased evaluation
Total	5,558	~3,736	Full 3-channel (grayscale replicated) acoustic imagery

---

🛰️ Sensor & Multi-Source Composition

1. NOAA Klein 5000 High-Resolution Surveys (unknown_debris):
   • Derived from public NOAA hydrographic surveys (H11833, H11835, H11836, etc.) using Klein 5000 side-scan sonar.
   • Covers complex seabed geologies: silt, sand ripples, rocky outcrops, and varying acoustic shadows.
2. MILCO Mine Countermeasures Klein 3500 Dataset (mine):
   • Spans MCM exercise surveys from 2010, 2015, 2017, 2018, and 2021.
   • v6 4-Point Calibration + train-only oversampling:
     • NOMBO Fusion: Co-occurring mine-like objects (NOMBO) mapped to master mine class to eliminate false-negative penalties.
     • Sub-pixel Box Cleansing: Removed corrupted bounding boxes (< 2.5 pixels).
     • Clutter Rebalance: Negative backgrounds capped to avoid probability dampening.
     • Stratified Multi-Year Partitioning: Balanced distribution of seabed topographies across years.
     • Plus train-only oversampling (airplane 3×, mine 2×) applied post-split.
3. Kaggle Side-Scan Sonar Benchmark (airplane & wreck):
   • High-fidelity acoustic returns of sunken vessels and aircraft fuselages.
   • Train-only augmentation (not synthetic images): 3×/2× noise-based copies of real training images (aspect ratio, shadow elongation, sonar speckle simulation).

---

📁 Archive Structure

The dataset archive sonarvision_multisource_v6.zip (924 MB compressed) unzips directly into standard Ultralytics YOLO layout:

sonarvision_multisource_v6/
├── dataset.yaml
├── images/
│   ├── train/        # 4,033 images (.png, .jpg)
│   ├── val/          # 563 images
│   └── test/         # 962 images
└── labels/
    ├── train/        # YOLO format text annotations (.txt)
    ├── val/
    └── test/

YOLO Annotation Format

Each label file contains lines formatted as: <class_id> <x_center> <y_center> <width> <height> (all coordinates normalized between 0.0 and 1.0).

---

🚀 Quick Start: Training with YOLOv8 / YOLOv8-ESI

1. Download & Extract from Hugging Face

from huggingface_hub import hf_hub_download
import zipfile

zip_path = hf_hub_download(
    repo_id="Dinoman1221/sonarvision-multisource-v6",
    filename="sonarvision_multisource_v6.zip",
    repo_type="dataset",
)

with zipfile.ZipFile(zip_path, 'r') as zip_ref:
    zip_ref.extractall("./datasets")

2. Train with Ultralytics

from ultralytics import YOLO

# Load model architecture
model = YOLO("yolov8n.pt")

# Train on v6 dataset
model.train(
    data="./datasets/sonarvision_multisource_v6/dataset.yaml",
    epochs=50,
    imgsz=256,
    batch=16,
    patience=15,
    name="sonarvision_v6_experiment"
)

---

🧩 Application Compatibility Note

This dataset card covers splits and archive layout only (unchanged). The full KADAL application repository also provides optional supporting features alongside the core workflow — file-ingest utilities, NMEA/KML export helpers, and a simulated live-waterfall demo view — documented in the GitHub README.md.

---

📜 Citation & License

The dataset is released under the Apache 2.0 License for academic, civil, and environmental research.

@dataset{kadal_dataset_v6,
  author = {Ashish S and Team KADAL},
  title = {KADAL Multi-Source Side-Scan Sonar Dataset (v6)},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://github.com/Dinoman67/sonarvision}
}