Tushar365/prithvi-2.0-flood-detection-sen1floods11

Prithvi-2.0 300M - Fine-tuned for Flood Detection (Sen1Floods11)

This model is a fine-tuned version of the Prithvi-2.0 300M foundation model, specialized for binary flood detection using Sentinel-2 optical imagery.

Model Description

• Developed by: Tushar Thokdar
• Model Type: Semantic Segmentation
• Backbone: Prithvi-2.0 300M (ViT-Base)
• Segmentation Head: UPerNet
• Input Resolution: 224x224
• Input Bands: 6 (Red, Green, Blue, Narrow NIR, SWIR 1, SWIR 2)
• Fine-tuned on: Sen1Floods11 dataset

Performance Metrics (Official Test Split)

Metrics derived from fine-tuning for 80 epochs on the official Sen1Floods11 test split.

Model Metrics

Metric	Fine-Tuned (80 Epochs)	Baseline	Gain (%)
Flood IoU	0.7196	0.1339	+437.3%
Flood F1	0.8370	0.2362	+254.3%
Flood Precision	0.8902	0.1638	+443.4%
Flood Recall	0.7897	0.4234	+86.5%
Mean IoU	0.8396	0.3953	+112.3%
Overall Accuracy	0.9633	0.6741	+42.9%

Per-Class Metrics (Fine-Tuned)

• No Flood IoU: 0.9595
• No Flood F1: 0.9793
• Flood IoU: 0.7196
• Flood F1: 0.8370

Training Configuration

• Epochs: 80
• Batch Size: 16
• Learning Rate: 5e-5
• Loss: Dice (0.5) + Focal (0.5)
• Data Splits: Official (252 Train / 89 Val / 90 Test)

Inference Performance

• Throughput: 20.66 samples/sec (NVIDIA T4)
• Inference Time (Avg): 0.048s per sample

Usage Instructions

To use this model with the godel-train library:

import torch
from godel_train.models.factory import ModelFactory

# 1. Initialize model with appropriate config
model = ModelFactory.segmentation(
    backbone="prithvi_eo_v2_300m",
    num_classes=2,
    checkpoint_path="pytorch_model.bin" # Local or downloaded file
)
model.eval()

# 2. Prepare sample input (Batch, 6 Bands, 224, 224)
# Bands: Red, Green, Blue, Narrow NIR, SWIR 1, SWIR 2
sample_input = torch.randn(1, 6, 224, 224)

# 3. Run Inference
with torch.no_grad():
    prediction = model(sample_input)
    # prediction shape: [1, 2, 224, 224] (Logits)

    mask = torch.argmax(prediction, dim=1)
    # mask shape: [1, 224, 224] (0: No Flood, 1: Flood)

Data and Credits

• Dataset: Sen1Floods11
• Fine-tuning: Performed by Tushar Thokdar
• Foundation Model: IBM/NASA Prithvi-2.0