Back to Projects
Production+437% vs Baseline

Prithvi-2.0 Flood Detection

Fine-tuned NASA-IBM Prithvi EO-2.0 vision transformer with UPerNet segmentation head on Sen1Floods11 benchmark, achieving SOTA-level flood mapping from Sentinel-2 imagery.

0.72 IoU
Flood Detection IoU
HuggingFace
Prithvi-2.0UPerNetPyTorchSentinel-2HuggingFace

Overview

Fine-tuned the NASA-IBM Prithvi EO-2.0 300M foundation model for binary flood segmentation, achieving 0.72 IoU and 83.7% F1 on the Sen1Floods11 benchmark dataset — a +437% improvement in Flood IoU over the foundation model baseline.

Technical Implementation

  • Fine-tuned Prithvi EO-2.0 vision transformer (ViT-Base) with UPerNet segmentation head
  • Trained for 80 epochs on 224x224 Sentinel-2 imagery (6 bands: RGB, NIR, SWIR1/2)
  • Implemented hybrid loss function (Dice 0.5 + Focal 0.5) optimized for imbalanced flood/non-flood classes
  • Inference throughput: 20.66 samples/sec on NVIDIA T4 (48ms per sample)

Key Metrics

Metric Score vs Baseline
Flood IoU 0.72 +437%
Flood F1 0.837 +254%
Flood Precision 0.890 +443%
Mean IoU 0.840 —
Overall Accuracy 96.3% —

What Makes This Special

The baseline Prithvi-2.0 model achieves only 0.14 IoU on flood detection out of the box. Through careful fine-tuning with a custom hybrid loss function and optimized training schedule, this model achieves near-production quality flood mapping — a critical capability for disaster response.

Published with complete documentation, inference code, and reproducible training pipeline on HuggingFace Hub.