VITO Remote Sensing | WorldCereal Team
Rather than running satellite data through a single-purpose classifier, VITO is exploring what happens when you take a step back: letting a foundation model compress what it sees into a general-purpose representation, then asking what that representation actually contains.
The conventional pipeline for Earth Observation analysis is straightforward in principle: take a satellite image, run a model trained for a specific task, get a result. It works. But it also means that every new question e.g. which fields are irrigated, which areas are at risk of crop failure, where built-up land is expanding, requires starting more or less from scratch.
Foundation models offer a different starting point. Instead of training directly for an output, these models first learn to compress satellite observations into dense numerical representations aka “embeddings” that capture the essential structure of the landscape without committing to any particular question. The idea is that a good embedding should be reusable across many downstream tasks.
VITO has been testing this in practice. Within the ESA Worldcereal project, VITO finetunes a geospatial foundation model, designed for multi-temporal Sentinel-1 and Sentinel-2 data. From this model, the team generated an embedding dataset covering the full extent of the United Kingdom. The dataset was originally produced in support of large-scale agricultural monitoring work, but the embeddings themselves contain considerably more information than any single crop map would capture.
Making the embedding space legible
Embeddings are, by nature, not directly interpretable as each pixel is represented by a vector of 128 numbers that don’t correspond to anything visible in the image. To make this layer useful for exploration and prototyping, the team built an interactive browser-based viewer that lets users work directly with the embedding space.
The viewer offers several modes. A Principal Component Analysis (PCA) projection lets users visualise the full embedding space in three dimensions, colour-coded by the principal components, giving an immediate sense of how the model organises the landscape. A supervised classification mode allows users to draw polygons on the map, label what they see e.g. wheat, built-up, other crops and than watch a cosine-similarity classifier update its predictions in real time across the visible tile. An unsupervised K-means mode segments the scene without any labels at all, surfacing structure that was latent in the embeddings.
The results are immediately readable. In the UK dataset, wheat fields, grassland, built-up areas, water bodies and other crops classes cluster into clearly separated regions of the embedding space. The K-means results also relate closely with visible field boundaries, and the 3D PCA view makes the separability more apparent than abstract.
Towards a reusable EO data layer
The UK dataset is one example of what becomes possible when embeddings are treated as a first-class data product rather than an intermediate artefact to be discarded. Storing them as Cloud-Optimised GeoTIFFs, indexed via a STAC catalogue, means they can be fetched on demand and plugged into new workflows without re-running the foundation model.
VITO is now exploring what this looks like at continental and ultimately global scale, producing embedding archives that could serve as a shared foundation layer for multiple teams and applications. The cost of generating good embeddings is paid once; the downstream benefit compounds across every task built on top of them.
For teams working on EO applications, whether in agricultural monitoring, land cover mapping, or environmental change detection, reusable embedding products could significantly lower the cost of developing new services. The demo viewer is one way to demonstrate that potential, putting the embedding space directly in the hands of analysts who want to understand what the model has learned.
Explore the interactive Embeddings Viewer: https://worldcereal.github.io/embeddings-viewer/
Images: Representative Images from the WorldCereal Embeddings Viewer (VITO, 2025).
