Learning general-purpose representations that capture spatial, temporal, and contextual signals across Earth.

Embeddings, explained simply

An embedding is a short list of numbers that acts like a fingerprint for something complex (an image, text, or a place). Similar fingerprints mean similar content—enabling fast search, grouping, and prediction.

Key papers and main takeaways

Earth Embeddings: Towards AI-centric Representations of our Planet

  • Introduces Earth embeddings as an AI-native representation layer for geospatial data (a reusable “location representation” across tasks).
  • Frames embeddings as a bridge between databases (retrieval/indexing) and models (generalization/interpolation) across modalities and scales.
  • Outlines a community roadmap: standardized embedding products, evaluation, and tooling to make geospatial ML more reusable and comparable.

SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery

  • Proposes contrastive pretraining that matches satellite image features with their geographic coordinates to learn a location encoder.
  • Produces general-purpose location embeddings that transfer across many downstream tasks and improve geographic generalization.
  • Shows that geolocalized EO imagery can act as scalable supervision for learning “place representations” without dense labels.

Geographic Location Encoding with Spherical Harmonics and Sinusoidal Representation Networks

  • Highlights why global location encoding is tricky: naïve coordinate embeddings can create spherical artifacts (notably near the poles).
  • Introduces a principled global encoder combining spherical harmonics (sphere-native basis) with sinusoidal representation networks (SIREN).
  • Demonstrates strong performance across benchmarks, motivating INRs/location encoders as a foundation for global Earth representations.

Measuring the Intrinsic Dimension of Earth Representations

  • Studies intrinsic dimension as a label-free lens on “how much information” Earth representations actually use (vs. their ambient vector size).
  • Finds intrinsic dimension is often much smaller than the embedding size and varies with resolution and training modality.
  • Shows intrinsic dimension can correlate with downstream performance and reveal spatial artifacts, supporting diagnostics and model selection.

Where we are heading

We build Earth embeddings to enable:

  • global retrieval (“find places like this”),
  • robust transfer across regions and sensors,
  • multimodal fusion (EO, climate, maps, text),
  • and interpretable representations with diagnostics that help scientific trust and use.