moritzrengert1/wildfire_global

A Globally Balanced Wildfire Satellite Dataset for VLM Bias Evaluation

Overview

This is a geographically balanced dataset of wildfire events captured by Sentinel-2 satellite imagery, designed to benchmark vision-language models (VLMs) on wildfire detection and burn severity estimation across diverse global regions.

Each sample consists of a paired pre-fire and post-fire RGB chip (224×224 px, 10 m GSD) with continuous dNBR ground truth derived from Sentinel-2 NIR/SWIR bands. The dataset is balanced across six continents and five ecosystem types to enable controlled analysis of geographic performance disparities in VLMs.

Dataset Structure

Each example contains:

Field	Description
pre_image_path	Pre-fire Sentinel-2 true-color RGB PNG
post_image_path	Post-fire Sentinel-2 true-color RGB PNG
dnbr_image_path	dNBR severity map (grayscale PNG)
kind	burn or neg (paired no-fire control)
severity3_label	Ordinal severity: 0=low, 1=moderate, 2=high
dnbr_mean	Continuous mean dNBR value (ground truth)
continent	One of six continents
country	Country of the event
ecosystem_group	forest, shrubland_chaparral, grassland_savanna, cropland, wetland
latitude_band	tropical, subtropical, temperate, boreal_polar
split	train / val / test (70/15/15, event-level)

GeoTIFF versions of all images are also provided.

Key Design Choices

• Continental balance: Equal target count per continent to avoid geographic skew.
• Ecosystem stratification: Samples are distributed across ecosystem types within each continent so geographic and biome effects can be disentangled.
• Paired burn/negative: Each burn event has a matched no-fire control chip from a nearby unburned location, verified fire-free via MODIS and Sentinel-2 dNBR.
• Tight temporal window: Post-fire images are captured 5–25 days after the burn event to minimize seasonal confounds.
• Quality filters: Minimum dNBR threshold (0.15) and burn pixel fraction (10%) ensure burns are visually distinguishable in RGB. Negatives must have dNBR ≤ 0.05.

Data Sources

• Burn detection: MODIS MCD64A1 Collection 6 (500 m monthly burned area)
• Imagery: Sentinel-2 L2A via Microsoft Planetary Computer
• Land cover: ESA WorldCover 10 m
• Boundaries: Natural Earth 110 m

Intended Use

This dataset is intended for:

• Benchmarking VLMs on wildfire detection and severity estimation from satellite RGB imagery
• Evaluating geographic and ecosystem-level performance disparities
• Studying the limitations of general-purpose VLMs on remote sensing tasks