Building Frictionless Geospatial AI: Making AlphaEarth Foundations Embeddings Accessible
This post is cross-posted on the Taylor Geospatial Engine Blog
If you’ve been watching the rapid growth of geospatial AI, you’ve definitely heard the buzz about embeddings—compact representations of satellite imagery that can capture complex patterns and relationships in satellite data. This summer, Google announced the AlphaEarth Foundations (AEF) model, which represents a step change for our entire community. Along with the model, Google released pre-computed global embeddings at 10-meter resolution, freely available and ready to use without expensive infrastructure or deep learning expertise.
Why This Matters to our Community
With Taylor Geospatial Engine’s (TGE) current initiative, Fields of The World (FTW), teams are testing multiple model architectures to predict field boundaries globally. We wanted to understand where AEF embeddings fit:
– Could they boost boundary-prediction accuracy? – Should they be used as direct inputs? – Or should they be used as features that enhance other data streams?
Embeddings are evolving fast, and part of TGE’s mission is to help the community innovate with less friction. We want to get these embeddings into the hands of researchers and practitioners as quickly as possible.
Creating a Frictionless Ecosystem
TGE is committed to lowering barriers for anyone exploring new geospatial approaches. We work in the open, and through our partnership with Radiant Earth, we ensure broad access to data via their open data publishing utility called Source Cooperative.
When AEF embeddings first launched, they were only available through Google Earth Engine (GEE). The FTW community—across academia, industry, and beyond—wanted access to the embeddings in cloud-native formats outside the GEE ecosystem, so we stepped in to make it happen.
A Collaborative Lift
With the community cheering us on, Jeff Albrecht (Principal Software Engineer at LGND AI) did the heavy technical lifting to move massive volumes of data (465TB to be more specific), working closely with Google to make the transfer possible. We then replicated the full dataset on Source Cooperative – now available as a collection of Cloud-Optimized GeoTIFFs at source.coop/tge-labs/aef for anyone to use.
“The value proposition of embeddings is lowering the barrier to entry to use geospatial data that informs decision-making. To achieve that, we need to spend a lot more time thinking up front about the sustainable data ecosystems that go into how we create, understand, and share embeddings. So that’s what I did,” says Jeff Albrecht.
Looking Ahead
We’re excited to explore what AEF embeddings might unlock for FTW’s global field boundary predictions and equally excited to apply this same approach to other emerging embeddings (such as Tessera).
No matter how the experiments turn out, we’ve already celebrated a meaningful win: the next team won’t have to figure out how to access this data.
Want to explore the AlphaEarth Foundations embeddings yourself? Visit source.coop/tge-labs/aef to get started.
Free the embeddings!
Our blog is open source. You can suggest edits on GitHub.