A spectrally grounded benchmark for vision-language models in Earth observation.
1,724 Sentinel-2 L2A tiles (~5 km squares) with 11 spectral indices computed
per tile and per 3x3 region, plus RGB previews.
Status: work in progress. This release contains imagery statistics only.
Question-answer pairs and evaluation splits are not yet published — do not use
for evaluation yet.
Fully reproducible: every tile is identified by its STAC item ID, so the whole
dataset… See the full description on the dataset page:
https://huggingface.co/datasets/riotu-lab/SpectralCheck-EO.
