Europe on an equal-area map, each square coloured by its AlphaEarth embedding

Europe, Unfolded

A tour through Europe's latent space.

Each dot is one 6 km square of Europe, 162,800 of them across 41 countries. Arrange them by their AlphaEarth Foundations embeddings, by terrain and land cover, or by where they are.

↓

Places

Which places in Europe look alike, which look like nowhere else, and the shortest walk between two of them.

The Alps, the Po plain and the Adriatic coast, each square coloured by its AlphaEarth embedding and cut at the coastline

Europe in 64 Dimensions

Every 6 km cell of Europe coloured by its embedding.

 Satellite views of the tour stops, the places with no near twin in Europe, numbered in order: a lava field on Etna, a salt lake, sand dunes, marsh channels, a glacier, salt pans

Europe's Most Unusual Places

The places that look like nowhere else, and the one place each resembles.

 A gold walk from Connemara to Crete over the embedding colours of Europe, by way of Wales, Cornwall, Brittany, the Spanish north coast, Corsica and the Maghreb, a dot at each stop

A Walk Through Latent Space

Pick two places in Europe and follow the chain of look-alikes between them.

 Sentinel-2 views of two distant mountain places the embedding finds alike, the distance between them marked

Doppelgängers

Places a continent apart that are each other's closest match.

Time

AlphaEarth publishes a new embedding every year. Comparing the years picks out where cities grow, which forests die back, when solar parks go up and how long a burn scar lasts.

Zurich at street scale in 2024, brighter where a place's texture that year was unlike its usual self

How Cities Change

Which places in a city looked unlike themselves in a given year, beside the satellite view of what was there.

 Year clocks for a steady place, a variable one, a trending one and one with an anomalous year

Where Europe Is Changing

Which places hold still, which had an anomalous year, and where the order of the years matters.

 Harz forest stands coloured by when they lost canopy, one ringed, with an inset of its embedding over time

The Year Before the Dieback

Which German forest stands went on to lose canopy, read from their 2018 embedding and checked against DLR's FCCL product.

 Sentinel-2 views of one German solar park, open fields, then rows of panels

Watching Germany Go Solar

When each German solar park was built, read from its embeddings alone and checked against the Marktstammdatenregister.

 Sentinel-2 views of one burnt German forest: the summer before the fire, a year on, and years later

How Long a Wildfire Is Remembered

Burnt ground told from its unburnt neighbour years after a European wildfire, scored on held-out fires against EFFIS burnt areas.

Geometry

How the 64-dimensional space is laid out: one direction follows summer heat, averaging shortens the vectors, and a decoding error once tilted every one of them.

Europe from Britain and Spain to the Balkans in summer: heat predicted from the embedding left of a gold line through the Alps, measured to its right, both with latitude and altitude taken out, blue where cooler than that leaves and orange where warmer, with country borders and the sea in dark grey.

What the Satellite Knows About Summer

How much of summer heat a straight line through the embedding recovers.

 An illustration: an eight-by-eight grid of small white arrows on green, tan and grey squares, one direction per land cover, with a faint dashed ring and a single gold arrow showing their average, which falls short of the ring.

Averaging Embeddings Pulls Them Toward Zero

Averaging an embedding over 6 km cells shortens it, most where the ground is mixed.

 A decoding curve with two codes on it and the straight line between them: the average of the two values, on the line, sits above the value of the averaged code, on the curve, under the equation decode(mean c) ≠ mean(decode c).

An Average Gotcha

The bug that turned every cell, and what it did not change.