Earth Embeddings as the Foundations of Geospatial Analysis

Yongchuan Cui , Keli Shi, Shunlin Liang

Earth observation satellites generate petabytes of imagery each year, but extracting useful information remains constrained by limited labels, heterogeneous sensors, and the cost of processing large archives. Foundation models reduce part of this burden by learning transferable representations from multi-source observations. More recently, these representations have been precomputed over continental and global extents and released as embedding layers or datasets, with AlphaEarth and TESSERA as representative examples. This survey examines Earth embeddings as foundations for geospatial analysis through three interdependent questions: Represent, Relate, and Reason (3R). This perspective connects how embeddings represent the world, reveal relationships, and support scientific understanding. We compare pretrained models, reusable encoders, and released embedding products, examining how their learning principles and design choices influence these capabilities. We also review access infrastructure, evaluation benchmarks, and applications to assess current capabilities and limitations. Experimentally, we compare the structure and predictive utility of different embeddings across geospatial tasks. Finally, we summarize the technical and scientific challenges and outline future directions for developing, evaluating, and using Earth embeddings.