In this thesis I investigate ship detection in satellite imagery using convolutional neural networks (CNNs). The goal is to develop a computer-vision system that automatically counts anchored vessels in optical satellite images and displays the occupancy of remote coves to boat owners.
I first compiled a three-part dataset from the Planet Explorer, xView and MASATI collections. The Planet Explorer subset contains 1 000 high-frequency scenes from the PlanetScope constellation, manually annotated in Label Studio with approximately 7 000 bounding boxes. I adapted the YOLOv11x model for small objects by freezing the initial convolutional layers, reducing mosaic augmentation and scaling, and training four variants—three single-source and one combined.
The results show that ships can be detected cost-effectively at a spatial resolution of 10 m/pixel, provided the objects exhibit sufficient contrast. For future work I propose incorporating synthetic-aperture radar (SAR) imagery to improve cloud robustness, ensembling multiple models, and expanding the dataset with high-resolution optical scenes. Together, these enhancements would turn the system into a reliable tool for real-time anchorage monitoring and broader maritime surveillance.
|
