Datasets:

---

Giezi

/

MicroWeather-Bench

Like 2

Adds the TESSERA version 1.1 embeddings for 2019 and 2024, next to the existing version 1.0 store (TESSERA_V1_ANNUAL_2024-CONUS.zarr), which is left untouched here:

• node-features/foundation-model-embeddings/TESSERA_V1_1_ANNUAL_2019-CONUS.zarr
• node-features/foundation-model-embeddings/TESSERA_V1_1_ANNUAL_2024-CONUS.zarr

What they are. For each of the 11,849 stations, the 128-d TESSERA 1.1 vector of the 10 m pixel containing the station, read from the published global map (dataset v1.1-dclimate, the complete global run of TESSERA 1.1 for 2017 to 2025, served from Source Cooperative / AWS Open Data) with geotessera 0.11.0 (GeoTesseraZarr.sample_points) on 2026-10-15. Same station order, layout, encoding and attribute set as the v1 store; source names the dataset and model version, history the store URL, package version and fetch date. The scripts, with a README, are in the Marmot repository under Embedding-Generation/TESSERA/; they follow the TESSERA fetch of EarthDataGFMEmbeddingGeneration, moved to the current API and dataset.

Coverage. 11,848 stations are read from their own pixel. One station, STN06825 (115.906 W, 47.475 N, Idaho), lies in a small unwritten hole of the v1.1 map in both years and gets the nearest written pixel, 20 to 30 m away. A status variable along station_id records this (valid or nearest_3px; water / nodata / outside would mark a station without a vector, whose row would be zeros as in the v1 store, but none occurs). The 35 stations that are all-zero rows in the v1 store, several of them lying exactly on 0.1-degree tile edges, are all valid in v1.1.

Relation to the v1 store. TESSERA 1.0 and 1.1 are different model runs and their embedding spaces do not correspond: the per-station cosine similarity between the v1.1 and v1 vectors for 2024 is centred on 0 (10th to 90th percentile -0.09 to 0.09), and the v1.1 vectors are smaller (mean absolute value 0.85 against 2.12). Within v1.1, 2019 and 2024 vectors of the same station agree well (median cosine similarity 0.92). The v1 store can be removed in a follow-up if the benchmark standardises on 1.1.

Verification. Values in each store equal the fetched tables; the per-array Zarr metadata is identical to the v1 store's apart from the added status variable; the sampling is deterministic (repeated reads, other batch compositions and per-point reads give identical vectors).

🤖 Generated with Claude Code