KAIST AI restores satellite views hidden by clouds
Clouds and darkness often hide exactly the satellite views that disaster responders and environmental scientists need. KAIST researchers led by Professor Munchurl Kim built GeoCore-9B, a foundation model trained from scratch on 10 million satellite images paired with descriptions, along with each image’s location and ground scale. Adapted with a lightweight fine-tuning method called LoRA, it restores cloud-covered images and turns hard-to-read radar scans into photo-like views. KAIST says it beat existing task-specific models on key cloud removal measures in tests on public datasets. The work was accepted to NeurIPS 2026, and the team plans to test it on more sensors and regions, with an eye on tracking floods and wildfires.