Researchers at KAIST in South Korea have created GeoCore-9B, an artificial intelligence model with about 9 billion parameters that reconstructs ground hidden by clouds in satellite images.
The model also converts radar images, which see through clouds, into images that look like ordinary photographs. It was trained from scratch on 10 million satellite image-text pairs from around the world, together with latitude, longitude and resolution data.
According to the team led by Professor Munchurl Kim, GeoCore-9B outperformed specialised cloud-removal models on brightness, colour and structural similarity measures. The work has been accepted for presentation at the NeurIPS 2026 international conference.
The technology could help assess disaster areas and track changes in forests, farmland and cities. The team also plans to apply it to predicting floods, wildfires and building damage.
