EcoFireBias

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… See the full description on the dataset page: global.

Source: Hugging Face Hub (moritzrengert1/wildfire_global). Metadata imported from the dataset’s Hub tags.