The goal of this project is to accurately segment flood-affected areas in satellite imagery using a UNet model with Batch Normalization for improved training stability and generalization.
- Model: UNet architecture with added Batch Normalization layers.
- Loss Function: Binary Cross-Entropy (BCE) loss for pixel-wise classification.
- Metrics:
- IoU (Intersection over Union)
- Dice Coefficient
- Data: 290 images with data augmentation (flips, rotations, scaling, etc.) to improve model robustness.
- Challenge: Hand-labeled masks with inaccuracies.
- Validation IoU: 0.7851
- Validation Dice: 0.8729
- https://github.com/divyansh44/Cosmolligance2k24/blob/main/cosmo-base.ipynb
- https://github.com/divyansh44/Cosmolligance2k24/blob/main/Flood-Segmentation-Project-Robust-Model-Development-2.pdf
Team Purple
Members: Geeth, Green, Divyansh