Google AlphaEarth Embedding Geometry Characterized for Environmental Reasoning

Google AlphaEarth has developed 64-dimensional embeddings that capture land surface data from 12.1 million samples across the Continental United States, spanning from 2017 to 2023. The geometry of these Earth observation model representations has been analyzed, uncovering a non-Euclidean structure with an effective dimensionality of 13.3 derived from the original 64 dimensions. The local intrinsic dimensionality is estimated at 10, and significant rotation occurs in tangent spaces, with 84% of locations exceeding 60 degrees. The mean absolute cosine for local-global alignment is 0.17, close to the random baseline of 0.125. Supervised linear probes reveal that concept directions rotate throughout the manifold, and an agentic system has been created to utilize this geometric insight for environmental reasoning, with implications for subsequent reasoning tasks being explored.