Flood Segmentation (AISeHack)

What this project does

This project detects flooded areas in satellite images using a deep learning model.

Each pixel in the image is classified as:

• 0 → Background
• 1 → Water
• 2 → Flood

---

How it works

• Uses satellite data (SAR + optical bands)
• Adds extra features (NDWI, MNDWI, NDVI, SAR difference) to help detect water better
• Uses 2 models:
  • UNet++ (EfficientNet-B5)
  • UNet (ResNet34)
• Combines both models using a weighted average (0.8 / 0.2)

---

Results

• Kaggle Score (Flood IoU): 0.2215

---

Files in this repo

• notebook.ipynb → training + inference
• requirements.txt → dependencies
• LICENSE → ANRF license

---

How to run

1. Install dependencies: pip install -r requirements.txt

2. Open the notebook: notebook.ipynb

3. Run all cells

---

Model weights

Download here:
https://www.kaggle.com/models/kushsaha/flood-segmentation-model-aisehack

---

Notebook

View here:
https://www.kaggle.com/code/saanvibeledenaga/aquasentinels-top-model-phase-2-aisehack

---

Notes

• Adding spectral indices helped improve performance a lot
• Ensemble performed better than a single model

---

License

ANRF Open License