Google has officially launched AlphaEarth Foundations, a PB-scale AI model for integrating satellite data, accompanied by a groundbreaking 64-dimensional Satellite Embedding Dataset. This innovation distills multi-year, multi-satellite observations into a single 10m × 10m pixel, compressing dozens of data sources – including satellite imagery, radar, elevation, and climate data – into unified 64-dimensional "information capsules". This fundamentally redefines geospatial analysis methodologies.

About AlphaEarth Foundations

AlphaEarth Foundations is a powerful AI model developed by Google DeepMind. It fuses multi-source data – including optical satellite imagery, radar scans, LiDAR 3D mapping, and climate simulations – into compact, unified digital representations via embedding technology. These embeddings achieve a 16× reduction in storage requirements compared to conventional AI systems while enabling higher computational efficiency and analytical precision. The model provides a more comprehensive and consistent depiction of Earth’s evolving systems, delivering enhanced decision-making support for critical areas including food security, deforestation monitoring, urban expansion tracking, and water resource management.

Dataset Overview

The Google Satellite Embedding dataset is a global, analysis-ready geospatial resource. Available annually since 2017 at 10-meter resolution, this pre-processed dataset covers terrestrial surfaces and shallow water zones – including intertidal areas, coral reef zones, inland waterways, and coastal waters. Each 10-meter pixel contains a 64-dimensional geospatial “embeddings” , encoding the temporal trajectory of surface conditions measured by diverse Earth observation instruments and datasets within that pixel and its surroundings during a single calendar year.

Official Dataset Link:
https://developers.google.com/earth-engine/datasets/catalog/GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL

Dataset Applications

Official use case demonstrations include:

  1. Similarity Search: Select any location on Earth (e.g., a specific farmland or forest type) and instantly identify and map all global locations with analogous surface and environmental conditions.
  2. Change Detection: Detect land changes and track dynamic processes – such as urban sprawl, wildfire impacts/recovery, and reservoir water fluctuations – by comparing embedding vectors for the same location across years.
  3. Automatic Clustering: Group pixels into distinct categories without pre-existing labels using clustering algorithms. This enables automatic differentiation of landscape features like forest types, soil classes, or urban development patterns.
  4. Smart Classification: Generate accurate maps using minimal training data. Achieve high-precision classification results (e.g., crop type mapping) with significantly fewer samples, reducing computation time and resource requirements.

Looking Forward

This breakthrough transcends technical advancement; it represents a paradigm shift in spatial cognition. By compressing Earth into a 64-dimensional framework, geospatial analysis evolves beyond basic feature recognition ("picture description") into the realm of interpreting spatio-temporal essence. The next generation of geospatial intelligence, powered by Earth Engine, has arrived.