Satellite Road Segmentation
A U-Net that extracts roads from satellite imagery
- Role
- Personal project
- Year
- 2025
- Built with
- Python · TensorFlow · Keras
Overview
A deep learning project that segments roads in satellite images. I implemented a U-Net and trained it on the DeepGlobe Road Extraction dataset, optimizing for Dice and IoU.
Model
The encoder has four blocks (32 to 256 filters), followed by a 512-filter bridge and a decoder with skip connections. It uses batch norm after every convolution, dropout of 0.3, transposed convolutions for upsampling and a sigmoid output. That comes to about 7.8M parameters.
Training
It trained on 5,603 images (256×256) and validated on 623, with Adam and binary cross-entropy. Checkpoints saved the best validation Dice, the learning rate dropped on plateaus, and training stopped early when progress stalled. The reported validation Dice is 0.67 and IoU is 0.51.
Tech Stack
- Python
- TensorFlow
- Keras
- NumPy
- Jupyter