🦎Breathing New Life into Raster Data with AI Embeddings and FME🦖

Geospatial imagery and other raster datasets often sit underutilized in archives, rich with potential insights that are hard to extract. Today, a class of AI models called embedding models is changing that. By using embeddings, numerical representations of data that capture its essential patterns, we can unlock information in current and legacy raster data that was previously inaccessible.

In the past, embeddings were mostly an internal step in classification models. For example, a neural network might learn features that help distinguish a cat from a dog, but those features were not typically used directly. In the geospatial domain, you usually still had to train a model on your own labelled data to get usable results.

With generic models like DINOv3, you can now compute embeddings directly on your own imagery, on your own machines. You can still fine-tune them if needed, but as we’ll see, you can already do a lot just by using the pretrained model “as is”. Of course, you can also build on top of them, this is why Meta call it a vision foundation model.

Image from https://ai.meta.com/dinov3/

With the flexibility of Safe Software’s FME (Feature Manipulation Engine) as a low-code integration platform, non-coders can harness these advanced AI capabilities. In this article, we’ll explain embeddings in simple terms and walk through three real use cases where combining an embedding model with FME gives raster datasets a new life. Each scenario is actionable, with data and processes available for readers to experiment with themselves (see the download link at the end).

What Are Embeddings?

In the context of machine learning, embeddings are a way to convert complex data (like an image or text) into a list of numbers (a feature vector) that captures meaningful characteristics of that data. You can think of an embedding as a compact summary or fingerprint of the original content.

For images, an embedding might be a vector of a few hundred or a thousand numbers that collectively represent things like shapes, textures and structures in the patch, but not in an explicit human-readable way. The key idea is that similar data will have similar embeddings. If two image patches both contain dense forest, an embedding model will generate vectors for them that are numerically close to each other. The model doesn’t just look at individual pixels, but also at patterns and context.

This technique compresses vast amounts of visual information into a smaller set of features that represent the data’s meaningful semantics. The power of embeddings is that they make it easy for computers to compare and search content. Instead of comparing raw pixel values (which is impractical), we compare the embeddings. This unlocks operations such as:

• searching images by content
• clustering and classifying scenes by similarity
• detecting anomalies

In short, embeddings turn rich, unstructured data (like raster pixels) into a structured form that we can index and analyze efficiently.

As reference, the same concept was used for text in our RAG article.

A New Life for Existing Raster Datasets

How do embedding models breathe new life into rasters?

Consider a large archive of aerial photos or satellite imagery collected over the years. Traditionally, to find something in these images, say all areas with water bodies or dense forest,, you’d have to manually analyze each image or train a model after labelling a lot of them. An embedding model changes the game by unlocking the content inside the images without extra work. By running each raster image (or even each tile or small patch) through the model, we obtain an embedding vector that represents that image’s content. These embeddings can then be used in multiple ways:

• Content-based search : find what you need, fast.
  Because similar images have similar embeddings, you can query the archive for images that match a certain example or pattern. For instance, given one image patch showing a lake, find all other patches with similar water textures. This was nearly impossible to do at scale before; embeddings make it feasible to build a “search engine” for imagery.
• Clustering and discovery : reveal hidden structure.
  By grouping patches by their embedding vectors, you might automatically discover categories in your dataset: urban, forest, water, bare soil, etc. This unsupervised discovery can reveal trends (like expanding built-up areas or deforestation) that give the dataset new relevance.
• Enhancing analyses with new features : combine with other data.
  Once images are converted to embeddings (or further distilled into categories or scores), you can cross-reference this information with other datasets to get new insights or fond anomalies.

FME + AI: Low-Code Integration for Everyone

You might be thinking: “This sounds complex , do I need to be a programmer or AI specialist to do this?” . Thankfully, no.

This is where FME comes in. FME is a powerful data integration platform known for its flexibility in handling spatial data, and it provides a no-code/low-code interface. You build workflows by connecting graphical components (transformers) rather than writing code. With FME, you can connect to virtually any data source or format, including large rasters, databases, web services, and now even AI models, all without writing a single line of code.

In our case, the embeddings are computed with DINOv3 and stored either as NumPy (.npy) files per tile or in a database table (DuckDB) with a vector column. FME’s strength lies in how you use them, both before and after the embedding step:

• Before:
  • Select objects or areas of interest.
  • Build patch grids aligned to your rasters.
  • Orchestrate tiling and parallel processing for efficient computation.
• After:
  • Join embedding results with vector layers or other attributes.
  • Apply filters and thresholds, run spatial joins, group by polygons.
  • Automate reporting, visualization and publication in FME Flow.

FME is famously format-agnostic and flexible. It has dozens of raster transformers and support for hundreds of data formats, meaning it can read your legacy data (GeoTIFF, JPEG2000, proprietary formats) and write out results in whatever form you need. Its no-code interface lowers the barrier so non-developers can create sophisticated data pipelines in an understandable, reliable way. Want to load embeddings into a spatial database, or compare them against a shapefile of land parcels? It’s a matter of adding the right transformers.

FME essentially acts as the glue that brings data and AI together: it can connect to databases or APIs, handle large volumes, and schedule or automate tasks in FME Flow for production. Providing a simple app or self-service workspace for end-users on top of these workflows is straightforward.

The result is that advanced image analysis isn’t confined to data scientists: planners, analysts, or any GIS professional can leverage it through FME.

Now, let’s look at three concrete use cases that demonstrate how an embedding model + FME workflow delivers value. Each scenario uses an existing raster dataset and combines it with other data in FME to solve a real problem.

Use Case 1 : Find All Solar Panels from One Example

Question

We have city-wide orthophotos and want to map solar panels. There is no labelled training dataset, and we don’t want to build and maintain a supervised model. Can we start from one known panel and find similar ones?

Idea

Because DINOv3 embeddings capture texture and structure, we can:

1. Pick one patch that clearly contains solar panels.
2. Look for all patches whose embedding is closest to that reference in the 1024-dimensional space.
3. Bring those patches back to map coordinates and visualize them.

No supervised training, just similarity search in embedding space.

Workflow in FME (conceptually)

1. Rebuild the patch grid

   • Use FME to generate a grid aligned with the raster tiling (cells of 8 m × 8 m, the physical size of a 16×16 pixel patch at 0.5 m resolution).
   • Each grid cell corresponds to a DINOv3 patch, identified by tile name and patch indices (row, column).

2. Select a reference patch

   • Load orthophotos as background in Data Inspector.
   • Click on a grid cell that clearly covers solar panels (good orientation, clear panels, limited shadow).
   • Extract its tile/row/col and look up its embedding in the DuckDB table.