We’re opening up Spheer FM embeddings.
Our first public release is now available on Hugging Face: 10 m, per-pixel, annual embeddings derived from Sentinel-2 time series.
For the initial release, we’re publishing:
• 10 MGRS tiles across the Netherlands
• 2017–2020
• 100-dimensional embeddings
• Cloud-Optimized GeoTIFFs
And we’re going to keep opening up more. In particular, we want to release the regions needed to run established benchmarks.
We’ve already benchmarked Spheer FM on TreeSAT-AI and PASTIS, where its specialized temporal representations perform very strongly. As we open up those regions, others will be able to rerun those benchmarks and explore the representations themselves.
And if you want to train models on top of these embeddings without writing a single line of code, you can do that in the Spheer app.
If there is a benchmark, region or ecological dataset you think we should prioritize next, let us know.
The embeddings are available here: https://huggingface.co/datasets/spheer/spheer-fm-embeddings
