Add the 2019 TerraMind embeddings (window and RGB-patch variants)
Adds the two TerraMind stores for 2019, the year the published Marmot model was trained on, next to the existing 2024 ones:
node-features/foundation-model-embeddings/TerraMind_v1_base_S2L2A_DEM_LULC_WINDOW_ANNUAL_2019-CONUS.zarrnode-features/foundation-model-embeddings/TerraMind_v1_base_S2_RGB_SEASONAL_PATCH_ANNUAL_2019-CONUS.zarr
What they are. The same two computations as the 2024 stores, run on 2019 inputs downloaded from Earth Engine with the same recipe (window variant: 12 monthly cloud-masked Sentinel-2 L2A composites, Copernicus DEM and ESRI land cover, 224 x 224 px at 10 m, centre 2 x 2 patches of terramind_v1_base, median over patches and months; RGB variant: true-colour 16 x 16 px crops of 4 seasonal composites, one token each, averaged). Same 11,849 stations in the same order, same layout, encoding and attribute set; temporal_coverage is 2019. The scripts that produce them, with a README, are in the Marmot repository under Embedding-Generation/TerraMind/.
Cloud-masked pixels are filled, not propagated. A pixel that is cloudy in every scene of a month (or season) is non-finite in the composite. The 2024 window store fed such pixels to the model unchanged, which turns the whole vector into NaN; those vectors were written as zeros (5,071 of 11,849 stations are all-zero rows in the 2024 window store). Here such pixels are set to TerraMind's band mean (0 after normalisation) before the model, the rule the CROMA extractor of EarthDataGFMEmbeddingGeneration uses for nodata, so every station has a vector. For stations without masked pixels the result is bit-identical to what the 2024 recipe gives (checked on 2024: all 6,778 such stations reproduce the Hub store exactly). Three per-station variables record what was filled so users can filter:
| variable | window store | RGB store |
|---|---|---|
masked_fraction |
share of the 12 x 12 x 224 x 224 pixel values filled | share of the 4 x 3 x 16 x 16 crop values filled |
masked_months / masked_seasons |
months with at least one filled pixel | seasons with at least one filled pixel |
blank_months / blank_seasons |
months filled everywhere (the model saw a flat image) | seasons filled everywhere |
In the 2019 window store 6,049 stations have at least one filled pixel and 1,393 have at least one blank month (1,146 one, 224 two, 21 three, 2 four); in the RGB store 13 stations have filled crop pixels (7 with a blank season). The history attribute names the filling step.
Verification. Values in each store equal the tables produced by the scripts; the per-array Zarr metadata is identical to the 2024 stores' apart from the three extra variables; the coordinates are the station coordinates of stations/locations.geojson. Computed on CPU, which reproduces the 2024 stores bit for bit (a GPU run differs at the 1e-6 level, and by up to 0.09 for a few stations).
Two caveats inherited from 2024. The land-cover modality receives ESRI classes remapped to 1..9 although TerraMind's LULC tokenizer was trained on the ESA WorldCover legend; kept for comparability. The RGB variant's brightness stretch previously turned masked pixels into black pixels without producing NaN; 11 stations of the 2024 RGB store are affected, and 67 further stations of that store predate a re-download of their seasonal chips. Regenerated 2024 stores with the filling policy are ready and can be proposed in a separate PR.
The dataset card (PR #1) will be updated to list these stores once the split presentation is settled.
🤖 Generated with Claude Code
Closed without merging: this PR could not be merged automatically (.gitattributes conflict with the PRs merged on 2026-10-17). Its content is carried unchanged, byte for byte, by PR #12, which consolidates every pending change on today's main.