This presentation introduces Google's AlphaEarth Foundations, an AI model designed to address the challenge of processing and analyzing petabytes of raw Earth observation data. Described as a "virtual satellite," the model ingests a vast array of multi-source, multimodal data—including optical, radar, LiDAR, climate, and text-based information—to generate a unified, analysis-ready digital representation of the planet's terrestrial and coastal surfaces.
The primary output is the Satellite Embedding dataset, an annual collection of 10-meter resolution images from 2017 to 2024, available in Google Earth Engine. In this dataset, each pixel is a precomputed, 64-dimensional embedding vector, creating a continuous, cloud-free data layer that eliminates the need for users to perform complex data preparation or run their own deep learning inference. The talk demonstrates how these embeddings can be used for a range of applications, including pixel-based similarity searches, unsupervised and supervised classification with significantly less training data, and robust change detection by comparing vectors across different years.
From Raw Pixels to Actionable Insights: Understanding the AlphaEarth Satellite Embedding Dataset
0:00 The Data Overload Problem: The session begins by highlighting the core challenge in Earth observation: while there is an unprecedented amount of satellite data from sources like Sentinel and Landsat, transforming these raw pixels into actionable insights is a major hurdle for scientists and analysts.
2:10 AlphaEarth Foundations: A Virtual Satellite: The AlphaEarth Foundations model is introduced as a solution. It's an AI model that assimilates disparate data streams (optical, radar, LiDAR, etc.) and compresses them into a single, information-rich, analysis-ready output called an "embedding." A key feature is its ability to create an "annual embedding," summarizing a full year of data into one cohesive image, something a physical satellite cannot do.
3:55 Multi-Source Data Ingestion: The model is pre-trained on one of the largest datasets ever created for this purpose, comprising over three billion image acquisitions. Its inputs include a mix of satellite imagery (Sentinel-1, Sentinel-2, Landsat), elevation models, climate information, and even weakly supervised text data from geolocated Wikipedia articles and species observations from GBIF.
6:35 The Satellite Embedding Dataset: The output of the model is a new dataset in Earth Engine. It consists of annual images where each 10-meter pixel is a 64-dimensional embedding vector. This creates a continuous, cloud-free surface, removing common issues like scan line gaps and the need for manual compositing.
8:48 Core Applications of Embeddings: The embeddings enable four key types of analysis:
Similarity Search: Find locations with similar environmental, spatial, and temporal characteristics.
Classification: Create highly accurate maps with far less training data than traditional methods.
Regression: Predict continuous variables like biomass or surface water percentage.
Change Detection: Identify changes over time by comparing embedding vectors from different years (the dataset spans 2017-2024).
15:16 How Similarity Search Works: A demonstration shows how to compare locations using the dot product of their embedding vectors. Because the vectors are unit length, the dot product directly calculates the cosine of the angle between them, providing a simple and efficient similarity score. An example shows how clicking on Manhattan can instantly highlight dense urban centers across the globe.
22:47 Simplified and More Accurate Classification: The embeddings serve as a very clean, low-noise feature space. This allows standard classifiers like Random Forest to achieve high accuracy with significantly fewer labeled data points. An example shows a crop classification map created with only 150 points per class that closely resembles the original high-density data layer.
26:31 Powerful Change Detection: Because the embedding space is consistent from year to year, comparing vectors from different years is an effective way to detect changes. A simple dot product between a pixel's 2022 and 2023 embeddings can reveal changes in land use or agricultural phenology.
39:54 Annual Data with Potential for More: The currently released dataset is annual, which is ideal for long-term monitoring but less so for short-duration events like floods. However, the model is capable of producing finer-granularity aggregations (monthly, weekly), and future updates will be guided by user demand and use cases.
55:17 No Need for Atmospheric Correction: The model is trained on top-of-atmosphere (TOA) data. The deep learning process is powerful enough to learn away atmospheric effects like clouds and haze, meaning users do not need to perform surface reflectance corrections themselves.