This presentation introduces Google's AlphaEarth Foundations model and its resulting Satellite Embedding dataset, designed to transform petabytes of raw Earth observation data into analysis-ready deep learning insights. AlphaEarth Foundations acts as a "virtual satellite," integrating diverse multimodal (optical, radar, LiDAR, climate, text) and spatio-temporal information into 64-dimensional embedding vectors for every 10-meter pixel. These precomputed, unit-length embeddings, hosted annually from 2017 in the Google Earth Engine catalog, are free of clouds and artifacts, significantly accelerating global mapping and monitoring. The presentation demonstrates the dataset's utility across common geospatial applications, including pixel-based similarity search, supervised and unsupervised classification with reduced annotation needs, regression, and robust change detection through time-series comparisons of embeddings. The initiative aims to provide a consistent, high-accuracy foundation for Earth analysis, with ongoing development based on user feedback.
AlphaEarth Foundations and the Satellite Embedding Dataset: Revolutionizing Earth Observation Analysis
0:31 Data Abundance & Challenge: The Earth Engine Data Catalog contains petabytes of satellite imagery, but transforming raw pixels into actionable insights is challenging due to data volume, complexity, and disparate sources.
1:39 The Solution: AlphaEarth Foundations: This deep learning model, acting like a "virtual satellite," assimilates vast amounts of Earth observation data (optical, radar, LiDAR, climate, gravity, text) into analysis-ready "information-rich embeddings."
2:23 Multimodal & Spatio-Temporal Integration: The model includes measurement context (optical, radar, LiDAR), spatial context (1.2 km field of view for each pixel's embedding), and temporal information (series of images to encode seasonal trajectories, phenology).
3:41 Annual Embeddings: Unlike real satellites, AlphaEarth Foundations can compress information and generate annual embeddings, providing a continuous monitoring record.
4:02 Diverse Data Sources: The model's pre-training dataset is one of the largest ever, with over three billion image acquisitions across five million sites, including contributions from Wikipedia and the Global Biodiversity Information Facility (GBIF) for weak supervision.
6:48 Satellite Embedding Dataset Structure: Each pixel in the dataset is a 64-dimensional embedding vector, forming a continuous 10-meter resolution surface without clouds, scan line gaps, or compositing issues.
7:10 Analysis-Ready Features: The embeddings are unit length (magnitude of one), simplifying machine learning calculations and making them "analysis-ready" by design, often easier to use than raw satellite images.
8:17 AI-Powered Analysis Foundation: The Satellite Embedding dataset is available in the Earth Engine catalog (collection ID provided) and serves as a foundational dataset for AI-powered analysis, eliminating the need for users to run their own inference or wrangle data.
9:00 Similarity Search: Compare an embedding vector from one location to others globally using the dot product to find places with similar surface conditions, environmental characteristics, or temporal patterns (e.g., identifying dense urban areas worldwide).
11:56 Unsupervised Classification (Clustering): Effectively explore the underlying structure in the embedding space using clustering algorithms (e.g., K-means) to differentiate landscape features.
22:54 Supervised Classification: Create highly accurate maps with fewer labeled training data points using standard Earth Engine classifiers (e.g., Random Forest, SVM) due to the low-noise feature space of embeddings.
9:48 Regression: Interpolate continuous variables like biomass or surface water across landscapes.
10:13 Change Detection: Identify temporal deviations by comparing embedding vectors for the same location across different years (available from 2017 to 2024), using metrics like the dot product (e.g., changes in agricultural phenology).
10:55 Tutorials Available: Five detailed tutorials are provided within the Earth Engine community to guide users on working with the dataset.
13:37 Linear Composability: Embeddings are linearly composable, preserving distance relationships during aggregation, making them well-suited for Earth Engine's pyramiding infrastructure.
23:33 Low Noise Feature Space: The embeddings represent a very clean signal, having already removed atmospheric effects, cloud cover, and other noise from original source data.
35:06 Global Terrestrial Coverage: The dataset covers Earth's terrestrial land surface and shallow nearshore waters, including coastal zones and small islands.
39:29 Encoder Model Availability: The AlphaEarth Foundations encoder model is not currently planned for direct public release.
40:00 Extreme Event Studies: While the annual embeddings capture temporal signals from events like floods, more frequent updates (monthly/weekly) are being explored based on user demand for near real-time applications.
42:11 Future Updates: Google plans to generate a new annual layer each year, with the 2025 layer expected in early 2026.
46:51 64 Dimensions Optimization: The 64-dimensional embedding space was chosen as a "sweet spot" to balance strong performance with usability in common remote sensing workflows and Earth Engine algorithms.