AlphaEarth Foundations is a new AI model that assimilates and encodes many different sources of geospatial information, effectively functioning like a virtual satellite. AlphaEarth Foundations can image the Earth’s terrestrial land surface and nearshore waters at any place and any time, generating a continuous embedding field that can be used as input for classification, regression, and change detection applications. In this talk, we introduce the embedding fields paradigm used to develop AlphaEarth Foundations. We’ll cover how we designed the model to handle data sparsity, multiple resolutions, and different bands and data types, as well as how we use different objectives to structure the AlphaEarth Foundations embedding space. We’ll also introduce the Satellite Embedding dataset, a collection of 10-meter-resolution annual embeddings for the years 2017 through 2024 generated by AlphaEarth Foundations. This analysis-ready collection of embeddings has over 1.4 trillion footprints per year, and can be used in place of any standard image inputs or other learned features for many different mapping applications and geospatial analysis workflows. We’ll share a few examples from our own validations and some applied use cases including land cover classification, crop type mapping, and disturbance monitoring.