title	Prithvi EO V2 — Flood Detection
emoji	🌊
colorFrom	blue
colorTo	indigo
sdk	docker
pinned	false
license	apache-2.0
app_port	7860

🌊 Prithvi EO V2 — Flood Detection

Binary flood segmentation from Sentinel-2 imagery using Prithvi EO V2 300M, fine-tuned on Sen1Floods11.

✨ Features

• Upload your own Sentinel-2 GeoTIFF (6-band or 13-band) or choose from 13 demo images
• Adjustable flood probability threshold and tile overlap
• RGB composite, flood probability heatmap, and overlay visualizations
• Download results as georeferenced GeoTIFF (flood mask + probability map)
• Model weights auto-downloaded from HuggingFace Hub on first run

📊 Performance

Metric	Score
Flood IoU	0.7196
Flood F1	0.8370
Accuracy	0.9633

🏗️ Architecture

• Backbone: Prithvi EO V2 300M (ViT encoder with 3D patch embedding)
• Head: UPerNet with Pyramid Pooling Module
• Input: 6 Sentinel-2 bands (B02, B03, B04, B8A, B11, B12) at 224×224
• Output: 2-class segmentation (no_flood / flood)

🚀 Run Locally

pip install -r requirements.txt
streamlit run streamlit_app.py

👤 Author

Tushar Thokdar

• HuggingFace
• GitHub