Satellites revisit the same ground, so it should be possible to predict what a later image will show. Most models forecast raw pixel values, which clouds, shadows and sensor noise corrupt, so much of their effort goes on noise rather than real change.
Instead, forecast the embedding: the compact set of numbers a pretrained foundation model, trained on masses of Earth imagery, uses to summarise a patch of ground. The authors' transformer, a network that weights past observations by relevance, has two branches: one gives a best guess with an uncertainty per number, the other a spread of plausible futures.
Forecasting embeddings halves cosine distance, a measure of how far apart two vectors point, from 0.20 to 0.10, and the forecasts also flag unusual change without labels.