Datasets:

---

Giezi

/

MicroWeather-Bench

Like 2

One PR with everything still pending, built on today's main so it merges without conflicts; it replaces the four open PRs #4, #5, #6 and #8, which I am closing. Later changes will be added here as further commits rather than as new PRs.

Contents

• node-features/foundation-model-embeddings/GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL_2019_30m-CONUS.zarr (from #4): the 2019 AlphaEarth embeddings read at 30 m, the vectors the published Marmot model was trained with.
• GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL_2024-CONUS.zarr renamed to GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL_2024_10m-CONUS.zarr with the 10 m scale in its metadata (from #5); the old path is deleted.
• TerraMind_v1_base_S2L2A_DEM_LULC_WINDOW_ANNUAL_2019-CONUS.zarr and TerraMind_v1_base_S2_RGB_SEASONAL_PATCH_ANNUAL_2019-CONUS.zarr (from #6): the 2019 TerraMind embeddings, cloud-masked pixels filled as in the 2024 stores of #7.
• TESSERA_V1_1_ANNUAL_2019-CONUS.zarr and TESSERA_V1_1_ANNUAL_2024-CONUS.zarr (from #8): the TESSERA 1.1 embeddings. Since #9 removed the 1.0 store, main currently has no TESSERA store; this restores one.
• README.md: the dataset card gains the two OlmoEarth 1.2 stores merged in #10 (store table, model table, how the vectors were computed, the masked-pixel variables, provenance, reference) and the updated totals (23 stores, 548 files, 21.96 GB). The card already described the four stores above.

The store files are byte-identical to those of the closed PRs (same content, same hashes); only the base commit changes. After this merges, main holds every store the card lists.

🤖 Generated with Claude Code

Hi Jonathan, one more commit on this PR (59dd7a1). This one touches two of the terrain stores, and it changes labels only, no data.

What I found. The aspect store's two band labels are the wrong way round: the band labelled Eastness holds northness, and the band labelled Northness holds eastness. I checked it over all 11,849 stations against the downslope direction computed from this dataset's own DEM patches: band 1 correlates 0.86 with the north component and 0.00 with the east component, band 2 the reverse, and flat ground holds (1, 0) = (cos 0, sin 0).

What the commit does. It renames the labels to match the values (band_name, long_name, name, one history sentence, and the card). The 22 MB data file is not touched, so a model trained on the store by band position is unaffected.

Two requests.

1. Please keep the values in their current order. Our end-to-end models were trained on it, and a store with the two bands exchanged would feed them the wrong channels without any error.
2. The labels presumably come from the script that generated the store. It is worth correcting there too, so that a regenerated or extended store comes out with the right names.

Also in the commit: the DEM store's title said "Slope"; it now says "DEM".