kittkyatkai/ShipFinder

How It Works

Hugging Face Space

URL: https://huggingface.co/spaces/kittkyatkai/ShipFinder

Repository Structure

├── app.py              # Streamlit application
├── processing.py       # Segmentation pipeline
├── metrics.py          # Diagnostic metrics and overlays
├── utils.py            # Image loading and helpers
├── sample_images/      # Built-in sample satellite images
├── requirements.txt    # Python dependencies
├── docs/
│   └── design_choices.md
└── README.md

Known Limitations

• —Resolution dependence. The app is tuned for 10m/pixel Sentinel-2 imagery. At higher or lower resolutions, the default parameters will need adjustment.
• —No spectral band selection. Real Sentinel-2 data has 13 spectral bands; the app works on RGB composites only, which limits water/land discrimination.
• —Simple thresholding. A global brightness threshold works well for dark open water but struggles with sun glint, cloud shadows, and bright harbours. Adaptive thresholding or k-means could improve robustness.
• —No wake analysis. Ship wakes are visible at 10m resolution and carry information about heading and speed, but the current pipeline does not exploit them.
• —Synthetic sample images. The bundled samples are procedurally generated to resemble satellite imagery. Real Sentinel-2 tiles would provide a more rigorous test.