🌧️ SaTformer — Space-Time Transformer for Precipitation Nowcasting

A Space-Time Transformer that forecasts cumulative precipitation (4 hours ahead) from 1 hour of satellite radiance imagery. The model outputs a probability distribution over rainfall bins, visualised below as a cumulative mass function (CMF).

Paper: A Space-Time Transformer for Precipitation Nowcasting — NeurIPS 2025 🏆 1st Place CUMSUM

Sample scenario

Select a pre-packaged satellite input sequence

Seed

Satellite input (1h — water vapour channel)

Ground-truth radar (4h ahead)

Predicted precipitation distribution (CMF)

Expected rainfall

Ground truth rainfall

Examples

Model: leharris3/satformer · Code: github.com/leharris3/satformer

The model processes 4 frames of 11-channel satellite radiance at 32×32 spatial resolution and predicts a categorical distribution over 64 precipitation bins. The CMF plot shows the cumulative probability — the area to the left of the ground-truth line is the under-prediction mass, and the area to the right is the over-prediction mass.