Neeraksh – SAR Oil Spill Detection
Deep learning model developed for the SlickTrace maritime oil-spill intelligence project.
- Architecture: U-Net with ResNet34 encoder
- Input: Two-channel Sentinel-1 VV/VH SAR imagery
- Preprocessing: Clip SAR dB to [-30, 0] and normalize to [0, 1]
- Training: 2,569 Oil, No-Oil and Lookalike scenes
- Best validation IoU: 69.42%
- Best validation Dice: 81.95%
- Output: Oil-spill segmentation mask at threshold 0.5
This repository hosts the trained PyTorch checkpoint. It does not provide a live inference API.