Tushar365/prithvi-burn-scar-dataset

Prithvi Burn Scar Severity Dataset

A curated dataset of Sentinel-2 multi-temporal satellite imagery chips for wildfire burn scar severity segmentation, pre-processed for the Prithvi EO 2.0 foundation model.

Author: Tushar Thokdar

Dataset Description

This dataset contains 108 image chips derived from Sentinel-2 L2A satellite imagery covering wildfire-affected regions. Each chip is a multi-temporal, multi-spectral array ready for semantic segmentation training with the NASA-IBM Prithvi EO 2.0 Vision Transformer.

Key Features

• Novel Delta Channel: Each chip includes a computed difference frame (Delta = Clip(Post - Pre, -1, 1)) that explicitly encodes spectral change magnitude
• Multi-temporal: 3 time steps per chip (Pre-fire, Post-fire, Delta)
• Multi-spectral: 6 Sentinel-2 bands per time step
• dNBR-derived labels: Ground truth severity classification from Differenced Normalized Burn Ratio

Data Format

Temporal Images (temporal_images/chip_*.npy)

• Shape: (3, 6, 224, 224) per chip
• Dtype: float32
• Dimensions:
  • Axis 0 — Temporal frame: [Pre-fire, Post-fire, Delta]
  • Axis 1 — Spectral band: [B2(Blue), B3(Green), B4(Red), B8A(NIR), B11(SWIR1), B12(SWIR2)]
  • Axis 2,3 — Spatial: 224 x 224 pixels at 20m resolution

Masks (masks/chip_*.npy)

• Shape: (224, 224) per chip
• Dtype: int64
• Values:

Value	Class	Description
0	Unburned	No fire damage detected
1	Low Severity	Minor vegetation damage
2	Moderate-Low	Partial canopy loss
3	Moderate-High	Significant vegetation loss
4	High Severity	Complete vegetation destruction
255	Ignore	Invalid/cloud/no-data pixels

Class Distribution

Class	Pixels	Percentage
Unburned	1,369,197	27.00%
Low Severity	771,342	15.21%
Moderate-Low	924,961	18.24%
Moderate-High	822,241	16.22%
High Severity	1,182,633	23.32%

Total labeled pixels: 5,070,374 across 108 chips

Data Source

• Satellite: Sentinel-2 L2A (ESA Copernicus)
• Resolution: 20 meters per pixel
• Area of Interest: Northern California wildfire region
• AOI Bounds: [-121.90, 39.80, -121.50, 40.30] (West, South, East, North)
• Pre-fire window: 2024-05-15 to 2024-06-30
• Post-fire window: 2024-09-15 to 2024-10-30
• Cloud thresholds: Pre=20%, Post=30%

Processing Pipeline

1. Google Earth Engine — Sentinel-2 cloud masking (SCL band), median compositing, 6-band extraction
2. dNBR Calculation — NBR = (NIR - SWIR2) / (NIR + SWIR2), dNBR = NBR_pre - NBR_post
3. Severity Classification — Thresholded dNBR values mapped to 5 severity classes
4. Reflectance Normalization — Raw DN scaled to [0, 1] surface reflectance
5. Delta Channel Generation — Delta = Clip(Post - Pre, -1.0, 1.0)
6. Tiling — 224x224 pixel chips with quality filtering (95% valid pixels, 1% minimum labeled pixels)

Usage

import numpy as np
from huggingface_hub import hf_hub_download

# Download a single chip
image_path = hf_hub_download(
    repo_id="Tushar365/prithvi-burn-scar-dataset",
    filename="temporal_images/chip_000000.npy",
    repo_type="dataset",
)
mask_path = hf_hub_download(
    repo_id="Tushar365/prithvi-burn-scar-dataset",
    filename="masks/chip_000000.npy",
    repo_type="dataset",
)

image = np.load(image_path)  # (3, 6, 224, 224)
mask = np.load(mask_path)    # (224, 224)

# RGB visualization (Pre-fire)
pre_rgb = image[0][[2, 1, 0]].transpose(1, 2, 0)  # B4, B3, B2
pre_rgb = np.clip(pre_rgb * 3.5, 0, 1)

Citation

If you use this dataset, please cite:

@misc{thokdar2024burnscardataset,
    title={Prithvi Burn Scar Severity Dataset with Delta Channel},
    author={Tushar Thokdar},
    year={2024},
    note={Sentinel-2 derived burn severity dataset for Prithvi EO 2.0}
}