Physically Interpretable AlphaEarth Foundation Model Embeddings Enable LLM-Based Land Surface Intelligence

Researchers analyzed 12.1 million satellite samples to prove Google's AlphaEarth foundation model creates physically interpretable embeddings. These 64-dimensional vectors map directly to 26 real-world environmental variables like temperature and elevation with up to 97% accuracy (R²=0.97). The team then built a Land Surface Intelligence system using these embeddings, allowing LLMs to answer natural language queries about the environment with high accuracy (score of 3.74/5 in evaluations).