GMI-AI/Road-segmentation-app

Road Segmentation from Aerial Images

This Space demonstrates road detection from aerial/satellite imagery using a U-Net deep learning model.

Model Details

• —Architecture: U-Net (CNN-based encoder-decoder with skip connections)
• —Training Dataset: Massachusetts Roads Dataset
• —Input: RGB aerial/satellite images (resized to 256×256)
• —Output: Binary segmentation mask (road vs. background)

Performance

• —IoU (Intersection over Union): ~0.52
• —F1 Score: ~0.63
• —Precision: ~0.70
• —Recall: ~0.60

How to Use

1. 1.Upload an aerial or satellite image
2. 2.Click "Submit"
3. 3.View the detected roads highlighted in red

The model works best with:

• —Clear aerial/satellite imagery
• —Visible roads (not heavily occluded)
• —Urban, suburban, or rural road networks

Limitations

• —Model trained specifically on Massachusetts aerial imagery
• —May not generalize well to very different geographic regions
• —Performance degrades with heavy occlusions (clouds, shadows)