AI · mapping

Finding similar features without training data

Point at a few example cells on a Sentinel-2 image; the foundation model ranks the other cells of the scene by similarity. The models are used as is, without labelling or training.

Expertise France · 2026

River meanders found by similarity search on a Sentinel-2 image
Tana valley, Sentinel-2 from January 2024, Clay model: four example cells on the river (green); the 40 best results (yellow) follow its meanders. Right: their thumbnails, to accept or reject.

Who it is for

For a first screening of an area (irrigated plots, settlements, quarries, water bodies), before finer mapping or to gather training examples.

What the application does

Search by example

A click on the map adds a positive or negative cell. Once embeddings are computed, the search updates on each click, validation or setting change.

Several models

Clay (80 m patch), TerraMind (160 m) and a no-AI baseline. Two models can be combined, for example optical and radar.

Validation and export

Accept or reject each result on its thumbnail, automatic threshold, linear classifier from 5 positives and 5 negatives. GeoTIFF and GeoJSON export.

How it works

  1. 01

    Scene

    Sentinel-2 L2A image (10 bands) over the drawn area, and optionally the Sentinel-1 image closest in time or a radar series.

  2. 02

    Embeddings

    One vector per patch, frozen weights: a few minutes per model on GPU for a scene.

  3. 03

    Score

    Cosine between each cell and the mean of the positives minus that of the negatives. A cell is the mean of the patches it contains.

Screenshots

Continuous score map of the similarity search
The same search as a continuous score map: how much each cell resembles the examples, across the scene.
Irrigated plots found by similarity search
Irrigated plots: eight example cells (green), 21 results above the threshold (yellow).

Known limitations

  • Cells of 80 m minimum (160 m for TerraMind): the tool finds areas, not small objects.
  • No quantified comparison yet between Clay, TerraMind and the no-AI baseline. For water or bare soil, the baseline may do as well.
  • Cloudy cells are excluded in optical imagery. Radar avoids this but measures something else: roughness and the structure of buildings and vegetation.

Going further

  • Count true positives on a few queries to compare the models
  • Export validated results as training examples

Role

  • Application design
  • Model comparison
  • Development and GPU testing

Technologies

PyTorch · TerraTorch · Clay · TerraMind · Sentinel-2 · Sentinel-1 · STAC · ipyleaflet · Voilà

Context

Demonstrator designed and built in 2026 by Guillaume Rieu, author of the Earth Observation Innovation Labs specifications, as part of the acceptance testing of these platforms in Kenya for Expertise France. It runs in the platform's JupyterHub environment.