This presentation introduces Google's Alpha Earth Foundations, a large-scale deep learning model designed to address the challenge of processing and analyzing vast amounts of Earth observation data. Acting as a "virtual satellite," the model ingests and synthesizes data from multiple sources—including optical (Sentinel, Landsat), radar, LiDAR, climate, and even geolocated text—into a unified, analysis-ready format. The primary output is the "Satellite Embedding Data Set," now available in Google Earth Engine. This dataset provides annual, 10-meter resolution, 64-dimensional embedding vectors that form a continuous, cloud-free, and noise-reduced representation of Earth's terrestrial surface from 2017 to 2024. The talk demonstrates key applications of these embeddings, including powerful similarity searches, highly efficient land cover classification with minimal training data, regression for continuous variables, and simplified change detection between years.
Unlocking Earth Data with Alpha Earth Foundations
0:00:39 - The Data Deluge: The presentation begins by highlighting the core problem: the massive and ever-increasing volume of free satellite data from programs like Sentinel and Landsat makes large-scale analysis difficult and time-consuming for individual users.
0:02:14 - A "Virtual Satellite" Solution: Alpha Earth Foundations is a deep learning model that acts as a "virtual satellite." It assimilates disparate data sources (optical, radar, LiDAR, etc.) and compresses them into a single, harmonious, and analysis-ready data product.
0:04:07 - Massive, Multi-Source Training: The model was trained on one of the largest geospatial datasets ever created, comprising over 3 billion individual image acquisitions. It uniquely incorporates not just imagery but also geolocated text from Wikipedia and species observations from GBIF for a richer understanding of the landscape.
0:06:54 - Understanding Embeddings: The model's output is an "embedding field image," where each 10-meter pixel is a 64-dimensional vector. This vector is a highly compressed, information-rich signature of a location's properties, including its spatial patterns and temporal behavior over a year.
0:07:15 - Analysis-Ready by Design: A key advantage of the embedding dataset is that it is a continuous, cloud-free surface. The vectors are pre-normalized to a unit length, simplifying machine learning workflows by eliminating the need for scaling or normalization.
0:08:17 - The Satellite Embedding Dataset: The primary product is a foundational dataset available in the Google Earth Engine catalog. It contains annual embedding layers covering Earth’s terrestrial land and shallow nearshore waters from 2017 to 2024.
0:09:00 - Four Core Applications: The talk focuses on four main applications enabled by the embeddings: similarity search, classification, regression, and change detection.
0:15:23 - Powerful Similarity Search: Users can find areas with similar characteristics anywhere in the world by comparing embedding vectors. A simple dot product calculation reveals similarity, allowing for tasks like finding all dense urban areas globally with a single click on Manhattan.
0:22:57 - Efficient Classification with Less Data: The clean, information-rich embeddings allow for highly accurate land cover classification (e.g., crop types) using standard classifiers like Random Forest, but with significantly less training data than required for raw satellite imagery.
0:26:33 - Simplified Change Detection: By comparing the embedding vectors for the same location between different years, users can easily and effectively detect changes on the landscape, such as shifts in agricultural phenology or deforestation.
0:30:37 - A Holistic, Not Interpretable, Signal: The 64 dimensions of the embedding vector are not individually interpretable (i.e., one
Error: Couldn't find 5133 axis is not "greenness" or "moisture"). They represent a complex, learned combination of all input data that should be treated as a single feature vector.
0:39:54 - Annual Cadence and Future Potential: While the current public dataset is aggregated annually, the model is capable of generating data at a higher frequency (e.g., monthly). Future releases of more frequent data will depend on user demand and identified applications.
1:11:01 - Getting Started: Google has launched a series of five tutorials within the Earth Engine community tutorials section, providing code and guidance to help users begin working with the satellite embedding dataset.