Create custom categorical maps from what you see on the globe—no code required.
If you already use Google Earth for your work or community projects, chances are you spend plenty of time doing visual inspection. You pan across the globe, zoom into a coastline or mountain valley, trace property boundaries, and inspect how neighborhoods, forests, or reservoirs are evolving over time. While the human eye is good at recognizing these spatial patterns, visual observations alone cannot easily be quantified, shared with non-technical stakeholders, or integrated into analyses to track changes over time.
Off-the-shelf global land cover maps can help, but they may be too generic for local decision-making. A global map might classify a region as “Forest” or “Agriculture,” when your project actually requires distinguishing native forest from commercial timber plantations, identifying specific crop types, or tracking specific invasive species. Until now, though, creating a custom categorical map may have required specialized expertise in satellite imagery analysis and machine learning code.
Today, we’re thrilled to dive into a new classification tool in Google Earth on the web, helping make the jump from visual inspection to custom categorical mapmaking much easier. You simply draw a boundary around your area of interest, define categorical classes, and label a few sample points. The tool trains a classifier on an underlying dataset of “Satellite Embeddings” (more on this below!) and generates a custom categorical map layer across your area of interest with zero coding required.
Turn local knowledge into actionable data layers
Consider a land manager needing to track invasive vegetation and unhealthy trees within a large area of management. Using the classify tool, the land manager can outline their area of interests, define classes of “healthy oak trees”, “unhealthy oak trees”, and “invasive trees”, and drop sample points indicating where they know some of those trees actually are. As points are added, a custom categorized map of their entire area is generated. Instead of spending time manually tracing polygons, needing to go out and physically survey the entire large area, or implementing complex machine learning code, the team can now prioritize specific areas to investigate further for management actions.
For formal reporting, you may want to run downstream quantitative analysis like a map accuracy assessment. To run this type of analysis, use the export feature to download your labeled sample points as a GeoJSON file so you can rebuild your map or run a formal accuracy assessment in Google Earth Engine. For step-by-step instructions on setting up your boundary and classes, visit our official documentation.
The technology under the hood
At the core of the tool is one of the world’s most cutting-edge geospatial technologies – the annual AlphaEarth Foundations Satellite Embedding dataset. The AlphaEarth Foundations model analyzes multiple types of satellite data like multi-spectral optical imagery and radar observations across an entire calendar year, compressing seasonal vegetation dynamics, moisture patterns, and surface textures into a unified representation called an “embedding” for every 10-meter patch of Earth.
When you use the classify tool, Google Earth Engine uses machine learning (specifically, a random forest classifier) to model the class of each pixel in the selected site based on those embeddings and the labeled sample points you provide. By combining global representations generated by AlphaEarth Foundations with Google Earth Engine's machine learning and your ground knowledge, the tool turns what used to require complex coding into a code-free, interactive workflow.
How building classification maps in Google Earth works under the hood...
Practical tips to get the best results
When you zoom in and place a sample point inside a cell it is applied to that 10-meter pixel.
Understanding how the underlying embeddings interact with your visual basemap is essential for building dependable classifications. AlphaEarth Foundations embeddings capture land surface and environmental properties at a 10-meter resolution. This is coarser than what you see when you zoom in close on Google Earth’s high-resolution basemap imagery. When you zoom in and place a sample point inside one of the cells in the yellow grid, it is applied to that 10-meter pixel. This means that your classified map will capture aggregate properties, not fine-grained spatial details.
To achieve high quality results, keep a few additional best practices in mind:
-
Capture the variation within each class. For example, if you have a single “water” class, place sample points in lakes, rivers, shallow coastal ocean, and deep ocean. This helps the classifier learn the full spectrum of each class.
-
Make sure your classes apply over the entire year. The annual satellite embeddings summarize an entire calendar year. You can use historical imagery to get a sense of what a sample point looks like over the year.
-
Maintain a balanced distribution of training sample points across your classes. If you provide thirty points for a forest class but only two points for a water body, the classifier can develop a mathematical tendency to predict forest in ambiguous areas.
Try classification in Google Earth today
The classify tool in Google Earth is available today as an experimental feature. To explore the feature, launch Google Earth on the web, open or create a project, and choose Tools > Classify. You can also consult our official documentation for comprehensive guidance on advanced settings, classification systems, and visualization tools.
