Earth Embeddings as the New Foundations for Geospatial Analysis: A Comprehensive Review

The rapid growth of Earth observation data, particularly from satellite constellations, airborne sensors, in situ monitoring networks, and geospatial platforms, has created unprecedented opportunities for understanding the Earth system. However, transforming these massive, heterogeneous, multi-scale, and multi-temporal data streams into actionable geospatial intelligence remains challenging. Recently, Earth embeddings, such as AlphaEarth and TESSERA, have emerged as a promising approach to representing locations, pixels, image patches, objects, regions, trajectories, or spatiotemporal processes as learned vectors. These representations compress complex Earth data while preserving semantic, spatial, temporal, and relational information useful for downstream analysis. This review examines the emerging field of Earth embeddings through a Represent, Relate, and Reason framework. We first define Earth embeddings and explain their importance for integrating diverse Earth observation and geospatial data. We then review data modalities, embedding units, learning strategies, model architectures, existing embedding products, evaluation methods, and geospatial applications. Comparative experiments complement this review by examining embedding structure, predictive utility, and sensitivity to dimensionality. We also discuss key challenges and future directions concerning generalization across regions and time, interpretability, uncertainty, benchmark design, physical consistency, bias, and responsible use. This review provides readers with a structured understanding of what Earth embeddings are, how they are generated and evaluated, and how they can provide a reusable foundation for geospatial analysis and scientific inquiry.