30,927 32×32 Sentinel-2 L2A multispectral patches (12 bands resampled to
10 m) across ten ESA WorldCover classes, selected so that the visible
bands are uninformative by construction: every patch sits in a region of
RGB-mean space dominated by patches of other classes, while its
NIR / red-edge / SWIR response stays class-informative.
The dataset is a controlled probe for one question: does a model actually
use the… See the full description on the dataset page:
https://huggingface.co/datasets/calebrob6/similar-but-different.
