🛰️ SkySeaLand Dataset (COCO Format)

The SkySeaLand Dataset is a comprehensive high-resolution satellite imagery collection designed for object detection, classification, and aerial surveillance analysis. This version is provided in the COCO (Common Objects in Context) JSON format, making it natively compatible with industry-standard frameworks such as Detectron2, MMDetection, YOLO (COCO-compatible), and the TensorFlow Object Detection API.

The dataset bridges terrestrial, maritime, and aerial domains, providing a unified benchmark for developing computer vision models capable of generalizing across complex real-world environments.

---

📚 Overview

• Total Images: 1,307
• Total Bounding Boxes: 19,103
• Annotation Format: COCO JSON (_annotations.coco.json)
• Classes: Airplane, Boat, Car, Ship
• Image Resolution: High-resolution aerial views
• Geographic Coverage: Asia, Europe, Russia, United States
• Scene Types: Airports, coastal areas, harbors, highways, marinas, offshore regions

---

📊 Dataset Split Summary

The dataset is organized into three subsets for standard machine learning workflows:

Subset	Percentage	Image Count	Annotation File
Train	80%	1,048	train/_annotations.coco.json
Validation	10%	132	valid/_annotations.coco.json
Test	10%	127	test/_annotations.coco.json
Total	100%	1,307

---

📊 Class Distribution

Class Name	Category ID	Object Count
✈️ Airplane	1	4,847
🚤 Boat	2	3,697
🚗 Car	3	6,932
🚢 Ship	4	3,627

---

🧾 Annotation Specification (COCO Format)

Annotations follow the official COCO JSON structure.

📦 Bounding Box Format

[xmin, ymin, width, height]

• Coordinates are absolute pixel values.
• area represents bounding box area in pixels².
• Each annotation includes image_id, category_id, and bbox.

📄 JSON Structure Example

{
  "images": [
    {"id": 0, "file_name": "image1.jpg", "width": 640, "height": 640}
  ],
  "annotations": [
    {"id": 0, "image_id": 0, "category_id": 1, "bbox": [100, 200, 50, 30], "area": 1500}
  ],
  "categories": [
    {"id": 1, "name": "airplane"},
    {"id": 2, "name": "boat"},
    {"id": 3, "name": "car"},
    {"id": 4, "name": "ship"}
  ]
}

---

🧰 Data Source and Tools

📡 Data Source

Satellite imagery was meticulously sourced from Google Earth Pro under research and educational guidelines.

The dataset is curated specifically for benchmarking hybrid detection models across aerial, maritime, and terrestrial environments.

🏷️ Annotation Tools

• CVAT (Computer Vision Annotation Tool): Used for primary image annotation and bounding box labeling.
• Roboflow: Used for annotation format conversion, preprocessing, and dataset validation.

---

🧠 Research Applications & Suggested Experiments

• Hybrid Detection: Evaluate model performance across land and sea simultaneously.
• Framework Benchmarking: Compare Faster R-CNN, RetinaNet, YOLO, and other COCO-compatible detectors.
• Domain Adaptation: Train on one geographic region and test on another.
• Edge Deployment: Optimize lightweight backbones for real-time aerial surveillance systems.
• Cross-Environment Generalization: Measure robustness in diverse environmental conditions.

---

🚀 Getting Started

Example: Load COCO Annotations in Python

from pycocotools.coco import COCO

coco = COCO("train/_annotations.coco.json")
categories = coco.loadCats(coco.getCatIds())
print(categories)

---

📂 Repository Structure

Ensure the directory structure is organized as follows:

.
├── README.md
├── train/
│   ├── _annotations.coco.json
│   └── (image files)
├── valid/
│   ├── _annotations.coco.json
│   └── (image files)
└── test/
    ├── _annotations.coco.json
    └── (image files)

---

📜 License

This dataset is released under the MIT License.
You are free to use, modify, and distribute it with proper attribution.

---

📝 Citation & Credit

If you utilize this dataset in research or commercial projects, please credit:

Author: Md Zahid Hasan Riad

---

⭐ If this dataset supports your research, consider giving the repository a star.