prithvi-burn-scar-model

Prithvi EO 2.0 - Burn Scar Severity Detection

Fine-tuned Prithvi EO 2.0 (IBM/NASA Geospatial Foundation Model) for pixel-level wildfire burn scar severity classification using Sentinel-2 multi-temporal satellite imagery.

Author: Tushar Thokdar

Model Description

This model performs semantic segmentation on multi-temporal Sentinel-2 satellite imagery to classify burn severity into 5 classes. It builds on the NASA-IBM Prithvi EO 2.0 Vision Transformer backbone with a UperNet decoder, fine-tuned using a novel Delta Channel Algorithm and a Freeze-then-Unfreeze training strategy.

Architecture

Input Specification

• Shape: (Batch, Time=3, Channels=6, Height=224, Width=224)
• Temporal frames: Pre-fire, Post-fire, Delta (Post - Pre, clipped to [-1, 1])
• Spectral bands: B2 (Blue), B3 (Green), B4 (Red), B8A (NIR), B11 (SWIR1), B12 (SWIR2)
• Normalization: Surface reflectance values in [0, 1]

Output Classes

Performance

Per-Class F1 Scores

Training Details

Strategy: Freeze-then-Unfreeze

• Stage 1 (Epochs 0-5): ViT backbone frozen, only UperNet decoder trained
• Stage 2 (Epochs 6+): Full model unfrozen for joint end-to-end optimization

Hyperparameters

Loss Function

Hybrid loss combining class-weighted Cross-Entropy and Dice Loss:

Loss = CrossEntropy(class_weights) + 0.8 * DiceLoss(classes=[1,2,3,4])

Class weights: [0.7406, 1.3147, 1.0963, 1.2333, 0.8575]

Key Innovation: Delta Channel

Standard approaches use 2-frame stacks (pre + post). This model uses a 3-frame stack that includes an explicit spectral difference:

Delta = Clip(Post-fire - Pre-fire, -1.0, 1.0)
Input = Stack([Pre-fire, Post-fire, Delta])  # Shape: (3, 6, 224, 224)

This provides the model with direct change magnitude information, significantly improving boundary detection.

Usage

Quick Inference

import torch
import numpy as np
from huggingface_hub import hf_hub_download

# Download checkpoint
ckpt_path = hf_hub_download(
    repo_id="Tushar365/prithvi-burn-scar-model",
    filename="prithvi-delta-best.ckpt",
)

# Model class must be available for load_from_checkpoint
from model import PrithviFireSegmentation

CLASS_WEIGHTS = torch.tensor([0.7406347, 1.3146889, 1.0963433, 1.233306, 0.8574721])

model = PrithviFireSegmentation.load_from_checkpoint(
    ckpt_path, class_weights=CLASS_WEIGHTS
)
model.eval()
model.cuda()

# Load a sample chip (3, 6, 224, 224)
image = np.load("chip_000000.npy")
input_tensor = torch.from_numpy(image).float().unsqueeze(0).cuda()

with torch.no_grad():
    logits = model(input_tensor)
    prediction = torch.argmax(logits, dim=1).squeeze(0).cpu().numpy()

# prediction shape: (224, 224) with values 0-4

Dependencies

torch>=2.1.0
pytorch-lightning==2.1.0
terratorch
segmentation-models-pytorch
numpy==1.26.4

Files in This Repository

Limitations

• Trained on a single wildfire event in Northern California; generalization to other regions/fire types needs validation
• 20m spatial resolution may miss fine-grained burn patterns
• dNBR-derived labels have inherent uncertainty at class boundaries
• Performance on cloudy/smoky imagery has not been evaluated

Citation

@misc{thokdar2024prithviburnscars,
    title={Fine-tuning Prithvi EO 2.0 for Burn Scar Severity Detection with Delta Channel Algorithm},
    author={Tushar Thokdar},
    year={2024},
    note={End-to-end wildfire damage assessment pipeline using NASA-IBM foundation model}
}

Acknowledgments

• Prithvi EO 2.0 by IBM and NASA for the foundation model
• ESA Copernicus for Sentinel-2 satellite imagery
• Google Earth Engine for data access and preprocessing