S2-100K (SatCLIP)

S2-100K A complete re-host of the S2-100K dataset used to train SatCLIP (Klemmer et al., 2023). 100,000 globally sampled multi-spectral Sentinel-2 L2A image patches (256×256 px, 12 bands, uint16), with longitude/latitude coordinates for each patch. Why this re-host? The original Microsoft-hosted blob ( is the canonical source. The only public mirror, davanstrien/satclip, is missing some patches (discussion). This repo is a… See the full description on the dataset page:

Source: Hugging Face Hub (torchgeo/s2-100k). Metadata imported from the dataset’s Hub tags.