Prithvi-2.0 Vision Model Achieves SOTA Flood Mapping from RGB
600M-parameter model adapts satellite pretraining to drone imagery with 95% accuracy
Deep Dive
Researchers adapt the Prithvi-EO-2.0-600M vision foundation model (combined with UPerNet decoder) for flood mapping from low-cost airborne RGB imagery. Fine-tuned on two datasets (BlessemFlood21, NeuenahrFlood), the model achieves state-of-the-art water segmentation. It transfers well to new flood events with minimal additional data, improving fastest and reaching almost the performance level of full training, indicating strong transfer capabilities.
Key Points
- Prithvi-2.0-UPN achieves state-of-the-art accuracy on both BlessemFlood21 and NeuenahrFlood RGB flood datasets
- Zero-shot transfer outperforms CNN baselines, and fine-tuning with just 10% of target data recovers near-full performance
- Model adapts a 600M-parameter satellite-pretrained ViT to centimeter-scale drone imagery with minimal data
Why It Matters
Enables rapid, low-cost flood mapping for emergency response without requiring massive labeled datasets per event.