With imagery and geo-embeddings, if you have an idea of the types of change you want to look for, it’s easier to find them. If you want to find deforestation, the LGND platform can search for the specific signature of this concept in the embedding space. Here is an example of what it looks like to search for recent deforestation in Oregon.

However, in many cases, you may want to detect change in general without having an explicit idea of what the change is. You might want to monitor a city over time to study high-level changes like urban sprawl or large infrastructure build-outs. In these cases, change detection and classification proves more complicated, but embeddings allow you to more easily detect, classify, and describe these changes vs traditional CV classifiers.

The LGND Change Analyzer

The LGND change analyzer is a demo we built on top of the LGND API to showcase how we solve this problem in our products like Studio. 

How it works

A user starts by selecting an area of interest (AOI) and two years of imagery to compare. The largest detected changes are then grouped into themes if relevant (ex. “agricultural field clearing”), or left as an individual change. 

In the example for Austin, TX metro area below, there are six themes highlighted, including “Earthworks and Grading” and “Crop Growth and Canopy”

From there, the analyst can move between the high-level view and the individual evidence. Each result shows before-and-after chips, a natural-language description from the change description endpoint of the LGND API. 

The goal is to make large-area change analysis feel less like searching for a known category and more like asking, “What changed here, and what kinds of change are happening?”

Behind the scenes

The LGND change analyzer exploits embeddings through simple mathematical and statistical operations. First, the program goes through each cell in the AOI and finds the arrow that leads from one embedding vector to another, which we call the delta embedding. The size of this delta embedding corresponds to the magnitude of the change. We only analyze significant changes which correspond to having delta embeddings above a certain length.

These delta embeddings are then grouped into collections that point in similar directions. These represent the themes our application surfaces. For example, if many cells show signs of deforestation, we should expect to see their delta embeddings pointing in a similar direction along a potential “forestation” axis. A number of tricks are applied in this clustering process to ensure themes are grouped accurately.

On an individual level, the changes in the before and after chips are described using the LGND API’s change description endpoint, and the changes are highlighted using Meta’s Segment Anything Model. This helps highlight changes an analyst may miss when comparing two chips.

Conclusion

The LGND Change Analyzer shows how embeddings can turn imagery into something we can compare, measure, and interpret at scale. Instead of only searching for a predefined type of change, we can look across an entire area and ask what has changed most, how significant those changes are, and whether they form meaningful patterns. The magnitude of the embedding shift helps quantify change, while the direction of that shift helps group similar changes together.

This makes embeddings especially powerful for geospatial analysis. They allow analysts to move from manual inspection or narrow category searches toward a more open-ended workflow: detect the strongest changes, understand the broader themes, and then inspect the underlying imagery for evidence. In practice, this means embeddings are not just a way to represent satellite images. They are a foundation for building tools that help us understand how the physical world is changing.

Now, see it for yourself — run a Change Analyzer report on an area you know, and see what surfaces that you didn't expect to find.