SkySeaLand Satellite Object Detection Dataset

This dataset was created to support a research idea that transportation-related objects in satellite imagery can be detected with steady accuracy when they are labeled with clear and consistent annotations. It focuses on four object classes: airplane, boat, car, and ship, and includes scenes from airports, highways, harbors, marinas, and coastal regions. These locations were selected to cover different backgrounds, object scales, and environmental conditions, which helps in studying model performance in realistic satellite settings.

Images were collected using Google Earth Pro under fair use and academic research guidelines. Candidate regions were explored manually and exported in high resolution through the built-in tools of the software. After collection, each image was reviewed for clarity, relevance, and presence of target objects. Files with heavy noise, strong cloud cover, or duplicate viewpoints were removed. When needed, basic preparation such as cropping and resizing was applied to keep the focus on relevant areas while maintaining visual quality.

Annotation was carried out using CVAT and Roboflow. Each object instance was marked with a bounding box and assigned one of the four class labels. A separate verification pass was performed to maintain consistent box placement and correct class assignment across the dataset. For this release, all final annotations are stored in a single COCO format JSON file that follows the standard object detection structure. The images are organized into separate folders for training, validation, and testing. A Roboflow project link is provided so that users can view the dataset online, apply their own preprocessing pipeline, and export the same annotations into formats such as YOLO or Pascal VOC if required.

The dataset can be interpreted by loading the COCO JSON file, reading the category identifiers, and mapping them to airplane, boat, car, and ship. Each annotation entry provides bounding box coordinates and class information that can be used directly with common computer vision libraries. Researchers can use this dataset to test model generalization across land and sea scenes, evaluate multi-class detection performance, study small object behavior, compare detection architectures, and explore transfer learning strategies in aerial imagery.