Foundation Model Embedding

Generate embeddings from an EO foundation model encoder. Outputs a multi-band raster (one band per embedding dimension) or GeoParquet with embedding vectors per chip. Supports Prithvi-EO-2.0, OLMo Earth, Clay, and other models registered in registry/models/. Preprocessing (band selection, normalization, chipping) is driven by the model profile.

ID	eofm_embed
Domain	ml
Category	inference
Type	raster-to-raster
Version	1.0.0

view yaml ◦view github ◦

Parameters

model: string · default prithvi-eo-2.0-300m · Foundation model name from registry/models/
device: select · default auto · Compute device (auto selects CUDA if available)
batch_size: integer · default 16 · Number of chips to process per inference batch
overlap: integer · default 32 · Overlap in pixels between adjacent chips
output_format: select · default auto · Output format for embeddings. "auto" selects GeoParquet for chip-level embeddings (CLS token) and COG for pixel-level (dense spatial). Per Cloud Native Geo best practices.
quantize: select · default none · Quantize embeddings to reduce storage. int8 gives 4x reduction with no downstream performance loss (per AlphaEarth findings).