Skip to content

EMBED-CD

Generate annual land change maps from satellite imagery RAPIDLY, at small and large-scale using QGIS. Draw a rectangle, pick two years, press one button — then teach it what the changes are by clicking a few of them.

Built on AlphaEarth embeddings | Expanding to others in future

AlphaEarth Foundations (Google / Google DeepMind) describes every 10 m pixel on Earth with 64 numbers summarising a year of optical, radar and elevation data. Global, annual, 2017-2025, CC BY 4.0, free to read.

Change is the dot product between a pixel's two embeddings, remapped so 0 means identical and 1 means opposite — equation 8 of the AlphaEarth paper. Each embedding covers a year, so it catches changes in behaviour, not just colour on one clear day.

Author's Note: I plan to add support for other embedding datasets in the future --only if they are deemed suitable for the platform.

Features

  • Draw and run. Draw a rectangle, pick two years, press one button.
  • Any size of area. Tiles are read and scored one at a time, so a large area costs no more memory to map than a small one. Coarser Detail is what makes a very large one quick.
  • Self-adjust threshold. The score is continuous, so the cutoff is symbology - instant and reversible.
  • Classify by clicking. Label a few objects and the rest follow. Model updates its fit in real-time! Fun!
  • Check against imagery. Self-validate outputs with Sentinel-2 mosaics, year by year. Be confident in the changes the model is producing.

Outputs

Output What it is
Change map A raster of continuous change score, 0–1, on an absolute scale.
Data coverage A companion layer saying why a pixel has no result: no tile, or a year missing.
Change objects A GeoPackage of polygons, each carrying its area, change statistics and the embedding of what it covers.
Classes Your labels and predictions on those objects, saved with the run and portable to another area.
GeoTIFF export The whole mosaic as a single file, for anything downstream.

Start here