This presentation introduces Google's AlphaEarth Foundations, an advanced AI model designed to address the challenge of processing and analyzing vast, complex Earth observation datasets. The speaker explains how the model functions as a "virtual satellite," integrating multi-source, multimodal data—including optical, radar, LiDAR, and textual information—to generate a unified, analysis-ready output: the Satellite Embedding dataset. This dataset, available in Google Earth Engine, consists of annual, cloud-free, 10-meter resolution images where each pixel is a 64-dimensional embedding vector. The session provides a hands-on overview of the dataset's primary applications, demonstrating how these embeddings enable pixel-based similarity searches, highly accurate mapping with less training data through supervised and unsupervised classification, and robust temporal change detection by comparing annual layers. The presentation concludes with a Q&A session that clarifies technical details about the model's architecture, data sources, the choice of 64 dimensions, and potential future developments.
Unlocking Earth Observation with AlphaEarth Foundations and AI-Powered Embeddings
0:31 The Challenge of Big Data: The session begins by highlighting the core problem: while we have petabytes of satellite imagery from sources like Sentinel and Landsat, processing this massive, complex, and often noisy (e.g., cloud-covered) data into actionable insights remains a significant hurdle for scientists and analysts.
2:10 The Solution: AlphaEarth Foundations: The speaker introduces AlphaEarth Foundations, an AI model that functions as a "virtual satellite." It ingests and assimilates disparate data sources to create a single, unified, and analysis-ready data representation.
4:02 Multi-Modal Training Data: The model is trained on one of the largest datasets ever created for this purpose, comprising over three billion image acquisitions. It integrates a wide variety of inputs, including optical imagery, radar, LiDAR, climate information, and even geolocated text from sources like Wikipedia and the Global Biodiversity Information Facility (GBIF).
6:42 Analysis-Ready Embedding Vectors: The output is a dataset where each 10-meter pixel is a 64-dimensional embedding vector, not a physical measurement. These vectors are information-rich, encoding spatial and temporal patterns, and are pre-processed to be cloud-free, continuous, and normalized (unit length), making them ideal for machine learning workflows.
8:17 Introducing the Satellite Embedding Dataset: The primary output is a publicly available dataset in the Google Earth Engine catalog. It provides annual embedding layers for the entire terrestrial land surface and coastal waters from 2017 to 2024.
15:23 Application: Similarity Search: A key application is finding locations with similar characteristics. By calculating the dot product between a selected pixel's embedding vector and all others, users can instantly generate a global map of similar places, such as identifying all major urban centers from a single point in New York City.
22:54 Application: Supervised & Unsupervised Classification: The embeddings enable the creation of highly accurate maps with significantly less labeled data. The data is optimized for both unsupervised clustering (to discover natural patterns) and supervised classifiers like Random Forest, greatly reducing the manual effort required for tasks like land cover mapping.
26:31 Application: Change Detection: Because the embeddings are consistent from year to year, users can easily perform change detection. By comparing the embedding vectors for the same pixel between two different years, it's simple to identify and quantify changes, such as shifts in agricultural phenology or urban expansion.
35:58 Future Directions and Availability: The dataset is currently exclusive to Google Earth Engine, with plans to release new annual layers. The team is actively seeking user feedback to guide future developments, including the potential for finer temporal resolutions (e.g., monthly or weekly) and making the data available in other formats.