Add the 2019 AlphaEarth embeddings used to train the published Marmot model
Adds node-features/foundation-model-embeddings/GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL_2019-CONUS.zarr: the AlphaEarth embeddings the published Marmot model was trained and evaluated with. The existing AlphaEarth store holds the 2024 annual embedding; this one holds the 2019 annual embedding at the same 11,849 stations, so the benchmark can reproduce the published model's input and compare years.
What it is. For each station, the 64-d AlphaEarth Foundations Satellite Embedding V1 (annual image for 2019) at the pixel containing the station, read from an Earth Engine export of the 2019 image at 30 m scale (512 x 512 pixel chips centred on each station). The stored vectors are identical to the table Marmot read during training (station_embeddings/backbone_center_embed_2019.npz, which also served as the target table), reordered to the repository's station order. All 11,849 stations have a value; no NaN.
Provenance check. The table was recomputed from the 30 m 2019 chips with the extraction script's exact method (pixel containing the station coordinates; NaN or values below -2 set to 0) and compared with the training table: all 11,849 x 64 values match exactly, and no value needed cleaning (no NaN, none below -2). The chips have a pixel spacing of 0.00027 degrees of latitude, that is 30 m.
Format. Same layout as the 2024 store: Zarr format 3, consolidated metadata, band_data (station_id, band, y, x) float32 with the same sharding and Zstd settings, band 1 to 64, x = y = 0, latitude (station_id, y) and longitude (station_id, x) float64, and the same attribute set (title, source, history, references, institution, spatial_coverage, temporal_coverage = 2019). The per-array Zarr metadata is identical to the 2024 store's except the three attributes that carry the year and the creation step. The coordinates are the station coordinates of stations/locations.geojson (the 2024 store's coordinates differ from those by up to 0.6 m).
The dataset card (PR #1) will be updated to list this store once the split presentation is settled.
Renamed the store to GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL_2019_30m-CONUS.zarr and added spatial_resolution: 30 m to its attributes (title and history updated accordingly). The data are unchanged. PR #5 does the same for the 2024 store, which is at 10 m.
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.