Skip to content

Model resources and Hugging Face

Find the source code, default checkpoint or embedding product, and usage guide for all 22 canonical adapters. The links below were checked on 2026-10-13. They point to upstream resources; Earth Embedding does not republish the weights.

Browse from Python

from earth_embedding import get_model_resources, list_model_resources

prithvi = get_model_resources("prithvi")
print(prithvi["weights_url"])  # Hugging Face model card for the default 100M-TL model

for model in list_model_resources(category="location"):
    print(model["model_id"], model["weights_kind"], model["weights_url"])

These helpers read bundled metadata only. They do not download weights, initialize providers, or run inference. Aliases such as alphaearth and esri work too. Call list_model_resources() without a category for all 22 adapters; categories are image, location, and precomputed. Each entry includes model_id, name, category, input, docs_url, source_url, weights_url, weights_kind, and notes.

Full resource directory

A means an image encoder, B a location encoder, and C a precomputed embedding product. Inputs describe the current adapter path, which can be narrower than the upstream model's capabilities. All adapters expose pooled and grid outputs; the meaning of a grid differs by model (tokens, coordinates, or product pixels). Read the linked guide before comparing them.

Model ID / guide Class Current adapter input Upstream source Weights or product
remoteclip A · image Sentinel-2 RGB Source Hugging Face model
satmae A · image Sentinel-2 RGB Source Hugging Face model
satmaepp A · image Sentinel-2 RGB Source Hugging Face model
satmaepp_s2_10b A · image Sentinel-2 10 bands Source Hugging Face model
scalemae A · image Sentinel-2 RGB + ground resolution Source Hugging Face model
wildsat A · image Sentinel-2 RGB Source official download
prithvi A · image Sentinel-2 6 bands + time/location Source Hugging Face model
terrafm A · image Sentinel-2 12 bands or Sentinel-1 VV/VH Source Hugging Face model
terramind A · image Sentinel-2 12 bands Source Hugging Face model
dofa A · image Multispectral imagery + wavelengths Source Hugging Face model
fomo A · image Sentinel-2 12 bands Source official download
thor A · image Sentinel-2 10 bands Source Hugging Face model
satvision A · image 14 calibrated TOA channels Source Hugging Face model
anysat A · image Sentinel-2 10-band time series Source Hugging Face model
galileo A · image Sentinel-2 10-band time series Source Hugging Face model
agrifm A · image Sentinel-2 10-band time series Source official download
satclip B · location Longitude/latitude only Source Hugging Face model
geoclip B · location Longitude/latitude only Source official download
esri_gle B · location Longitude/latitude only Source official download
tessera C · product Published GeoTessera embedding tiles Source official product
gse C · product Annual Google Satellite Embedding product Source official product
copernicus C · product 0.25-degree embedding product Source Hugging Face dataset

Which artifact am I opening?

  • Hugging Face model links identify the checkpoint repository selected by the current adapter default. The Prithvi default is 100M-TL and the TerraMind default is 1.0 small. Environment overrides can select a different source; see the model guide and describe_model(...).
  • MVRL checkpoints are the rshf-compatible packages used for RemoteCLIP, SatMAE RGB, SatMAE++ RGB, and Scale-MAE. Their upstream author repositories are linked separately; these are not Earth Embedding publications.
  • Hugging Face dataset is already-computed data. The Copernicus adapter reads TorchGeo's GeoTIFF redistribution; the source link leads to the author's Copernicus-Embed dataset. It is not an inference-model checkpoint.
  • Official download links lead to the authors' release instructions. WildSAT uses Google Drive, FoMo uses Dropbox, and AgriFM provides GLASS/OneDrive links. GeoCLIP uses its author's Python package; Esri GLE uses the ArcGIS Living Atlas model item. A Hugging Face account is not a universal requirement.
  • Official product links describe published embeddings. gse retrieves Google Satellite Embedding/AlphaEarth output; it does not load an AlphaEarth encoder. tessera queries GeoTessera tiles; retain the dataset version and variant when reusing embeddings.

Each upstream model or dataset keeps its own license, citation, access terms, and preprocessing requirements. Follow its model card or release documentation. The directory is a discovery aid, not a lockfile or a claim that all sources have the same license.

Location encoders, including Esri

satclip, geoclip, and esri_gle encode coordinates without fetching an image. Earth Embedding accepts longitude, latitude consistently and converts to the underlying model's coordinate order. Their outputs are time invariant, so changing temporal does not create new observations. A coordinate grid is a sampled location prior, not satellite imagery at that date.

The Esri adapter downloads the official Global Location Encoder package and runs its location weights directly with PyTorch. See the Esri guide for cache controls and model limits. The GeoCLIP guide covers its optional location extra; the SatCLIP guide covers its checkpoint variants.

The library's 22 adapters include the two software extensions GeoCLIP and Esri GLE. The manuscript benchmark remains the frozen 20 systems (A: 16, B: SatCLIP only, C: 3); listing an adapter does not add a benchmark result. See validation status for the evidence behind each integration, and model selection for input and output choices.