Google Launches Custom Satellite Embeddings for On-Demand Geospatial Monitoring
Moving beyond annual datasets, the new AlphaEarth-powered tool enables near-real-time change detection for logistics and land management.
Google has announced the private preview of Custom Satellite Embeddings, a tool designed to provide high-frequency, on-demand geospatial data for organizations requiring more than annual updates. Last updated on February 26, 2025, by the Google Maps Platform team, this new offering leverages the AlphaEarth Foundations model to transform complex remote sensing data into analysis-ready representations. This shift allows developers to move from static yearly mapping to active monitoring of specific regions at intervals as frequent as every five days.
Previously, organizations using Google’s geospatial tools were often limited to the Satellite Embedding dataset, which provided a single snapshot for each calendar year. While useful for long-term trends, this annual cadence proved insufficient for fast-moving industries like agriculture or disaster response. The new custom capability effectively acts as a synthetic observer, merging optical imagery, radar signals, and elevation data into a unified 64-band representation that can see through cloud cover.
How does the AlphaEarth model simplify geospatial analysis?
The core innovation behind Custom Satellite Embeddings is the AlphaEarth Foundations model developed by Google DeepMind. Traditionally, a 12-location HVAC operator or a regional utility provider would need a team of specialized machine learning experts to process raw satellite pixels, handle cloud masking, and align disparate sensor data. We observe that this new framework removes those manual preprocessing bottlenecks by delivering "embeddings"—mathematical representations of physical space that are already optimized for analysis.
By condensing multiple data sources into these embeddings, Google allows businesses to bypass the high compute costs associated with raw image processing. For example, a dental practice in Leeds would have little use for this, but a national forestry agency could use these embeddings to track deforestation without maintaining a massive local server infrastructure. The model essentially translates the complexity of the physical world into a standardized digital language that standard software can interpret.
Shifting from static mapping to Custom Satellite Embeddings monitoring
The transition to a sub-annual frequency marks a significant change in how geospatial intelligence is consumed. Unlike the previous workflow where data was processed in large, infrequent batches, Custom Satellite Embeddings allow for rolling updates tailored to specific timeframes. This is particularly relevant for tracking crop growth cycles in regions like California's Imperial Valley, where monthly sequences can capture the exact window of harvest.
In comparison to traditional satellite imagery which requires human analysts to manually spot differences between two photos, these embeddings allow for automated change detection. By using mathematical dot product similarity between two embedding states, systems can instantly isolate changes, such as identifying the precise burn scar after a wildfire in Los Angeles. This level of automation was previously cost-prohibitive for most mid-sized enterprises.
Applications for logistics and utility infrastructure
For businesses managing physical assets, the ability to request data for non-calendar intervals is a critical update. A utility company monitoring vegetation growth near power lines can now trigger alerts based on biweekly updates rather than waiting for an annual report. Similarly, retail supply chains can use these tools to monitor climate-driven risks and ensure compliance with international deforestation regulations in near real-time.
We see this as a strategic bridge between low-resolution global monitoring and high-resolution, expensive bespoke aerial photography. While it does not replace the need for 50cm resolution imagery for minute details, it provides the environmental context necessary to direct those more expensive resources where they are needed most.
What this means for local businesses
For operators and agencies managing large-scale physical assets or environmental data, this update changes the feasibility of real-time monitoring. Here is how organizations should respond:
- Audit current monitoring cadences: Evaluate if your current quarterly or annual data refreshes are causing delayed reactions to environmental changes.
- Shift to embedding-based workflows: Explore migrating from raw image processing to embeddings to reduce compute overhead and development time.
- Request early access: Organizations in agriculture, public sector, and logistics should join the Enterprise Private Preview to test the 5-day frequency capabilities for their specific regions of interest.
Sources
Frequently asked questions
- What is the difference between annual embeddings and Custom Satellite Embeddings?
- The annual Satellite Embedding dataset provides a single, unified view of the planet for each calendar year, which is useful for historical analysis and long-term trends. Custom Satellite Embeddings, currently in private preview, allow organizations to request data for specific regions and custom timeframes, such as weekly or monthly. This allows for 'active monitoring' rather than just static mapping, which is essential for industries like agriculture that follow seasonal rather than calendar cycles.
- How does AlphaEarth handle cloud cover in satellite imagery?
- AlphaEarth Foundations acts as a 'virtual satellite' by synthesizing multiple data types, including radar (which can penetrate clouds) and LiDAR, alongside optical imagery. By combining these different sensors into a single 64-band embedding, the model can produce a consistent and unobstructed view of the Earth’s surface even in regions that are persistently cloudy, ensuring that monitoring is not interrupted by weather conditions.
- Do I need machine learning expertise to use these embeddings?
- One of the primary goals of this product is to reduce the need for specialized ML expertise. Because the data is delivered as 'analysis-ready' embeddings, the complex work of preprocessing, data cleaning, and sensor fusion is handled by Google. Organizations can use these mathematical representations to build tools for change detection and classification using standard data science workflows, significantly lowering the barrier to entry for geospatial intelligence.
Related reading
