Add the OlmoEarth v1.2 Base embeddings for 2019 and 2024
Adds OlmoEarth version 1.2 (Base) embeddings for 2019 and 2024, a new embedding source for the benchmark:
node-features/foundation-model-embeddings/OlmoEarth_v1_2_base_S2L2A_WINDOW_ANNUAL_2019-CONUS.zarrnode-features/foundation-model-embeddings/OlmoEarth_v1_2_base_S2L2A_WINDOW_ANNUAL_2024-CONUS.zarr
What they are. For each of the 11,849 stations, a 768-d vector from the OlmoEarth v1.2 Base encoder (Allen Institute for AI; Hugging Face allenai/OlmoEarth-v1_2-Base, commit 2e99a73, package olmoearth-pretrain 0.1.2), computed on the 12 cloud-masked monthly Sentinel-2 L2A composites of the year that also feed the TerraMind window store. The model sees the centre 48 × 48 pixels (480 m) of the 12 composites in one pass over the year (patch size 4: a 12 × 12 grid of patches per month, 1,728 tokens, the extent and token count the model was pretrained on). Cloud-masked pixels are handled by the model's own mask: a patch of a month that contains a masked pixel is marked MISSING and the encoder drops that token before attention, so masked pixels never enter the computation. The station vector is the mean of the last-layer tokens of the 2 × 2 patches that meet at the station (80 m × 80 m) over the 12 months, visible tokens only. Float32 on GPU; the result is bit-identical between runs and between the V100 and RTX 8000 GPUs used. Same station order, layout and encoding as the other node-feature stores; source names the model and its commit, history the steps, package versions, GPU and date. The scripts, with a README, are in the Marmot repository under Embedding-Generation/OlmoEarth/; they follow the OlmoEarth time-series extractor of EarthDataGFMEmbeddingGeneration, with the monthly composites as the time series, the pretraining window size, the mask passed to the model, and one vector per station.
Per-station variables along station_id, next to band_data: masked_fraction (fraction of the 48 × 48 × 12 bands × 12 months values that were cloud-masked), masked_months (months with at least one masked pixel in the window), blank_months (months masked everywhere in the window, which contribute no token) and pooled_tokens (number of tokens averaged into the vector, at most 48; more would mean no centre token was visible and the whole window was averaged). Every station has a vector in both years: no station lacked a visible token and none needed the whole-window fallback. Stations with at least one cloud-masked pixel in the 480 m window: 3,935 in 2019 and 3,269 in 2024 (the 2.24 km TerraMind window has 5,071 in 2024); with at least one fully clouded month: 2,043 and 1,729; vectors averaged from fewer than 48 tokens: 2,976 and 2,532 (the smallest: 16 tokens in 2019, 4 tokens in 2024 for one station with ten clouded months). Between the two years the vectors of the same station agree well (median cosine similarity 0.97, 10th to 90th percentile 0.94 to 0.98, against 0.66 for random station pairs).
Relation to the other stores. Different model, different input extent: the TerraMind window store uses a 2.24 km window with DEM and land cover and fills masked pixels with the band mean; this store uses the 480 m window the model was trained on and masks instead of filling. OlmoEarth 1.0, 1.1 and 1.2 are different model runs whose embedding spaces do not correspond; only 1.2 is used.
Verification. Values in each store equal the computed tables; the per-array Zarr metadata is identical to the TerraMind window store's apart from the four added variables; the root attribute set is the same.
🤖 Generated with Claude Code