PEAK* East Asia

~40×The area of Korea’s radar coverage

Built PEAK* datasetSatellite, radar, gauges, terrain

This presentation Rainfall retrievalSatellite imagery to rainfall

In development NowcastingRainfall hours ahead

In development DownscalingSharper local detail

S2RAIN: High-resolution rainfall maps across East Asia

IMERG target

Model prediction

Placeholder

Step 0 of 31

Neural network

One small step

From noise toward rainfall, guided by the satellite inputs

Straight paths from noise to rainfall

All variablesEvery satellite input, plus terrain

Without terrainTerrain switched off

Infrared onlyA single infrared channel

CSIEvent detection

FSSSpatial agreement

SSIMSpatial structure

Geostationary satellites reach beyond regional radar limits.

Multimodal inputs improve satellite rainfall estimation, a promising direction for nowcasting.

A large-scale dataset across East Asia sets up what comes next: nowcasting and downscaling.

AI4Sci Korea 2026 · Seoul

S2RAIN: Satellite-to-RainfallHybrid Generative Modeling for Precipitation Nowcasting

Rogelio Ruzcko Tobias^1,2

Minki Choo², Jungho Im², Yeji Choi¹

¹DI Lab, Inc. · ²Ulsan National Institute of Science and Technology (UNIST)

ruzcko@dilab.kr

AI4Sci Korea 2026

Thank you

Rogelio Ruzcko Tobias^1,2

Minki Choo², Jungho Im², Yeji Choi¹

¹DI Lab, Inc. · ²Ulsan National Institute of Science and Technology (UNIST)

Loading…