KAIST AI Restores Cloud-Obscured Terrain Images With Photorealistic Detail
Researchers have developed an artificial intelligence (AI) that infers terrain features hidden by clouds in satellite imagery and restores them to photorealistic quality. A team led by Professor Moon-Chul Kim of KAIST's School of Electrical Engineering announced on the 20th that it had developed 'GeoCore-9B,' an AI broadly applicable to the generation, restoration, and transformation of satellite imagery.
Satellite imagery is essential for assessing wildfire and flood damage and monitoring changes in forests and cities, but its use has been limited by shooting conditions and sensor characteristics. Optical satellites, which capture sunlight reflected off the Earth's surface, struggle to observe the ground when clouds are thick or at night. Synthetic aperture radar (SAR), which emits radio waves and analyzes the returning signals, can observe at night or on cloudy days, but its images differ in brightness and texture from ordinary photographs, making interpretation difficult.
To overcome the limitations of both optical and radar data within a single model, the research team built a common foundation for processing such data together. The core generative model was trained on 10 million pairs of global satellite images and descriptive captions, adjusting roughly 9 billion parameters during training to learn how to read terrain features. It is a foundation model structure applicable across multiple tasks: generation, restoration, and transformation. Notably, instead of borrowing trained weights from general photo-generation models, it was trained from scratch on satellite imagery.
The team also incorporated latitude, longitude, and ground sample distance (GSD) into the image captions so that the AI reflects both the shooting location and real-world spatial scale. With a GSD of 1 meter, each pixel corresponds to an area of roughly 1 by 1 meter, allowing the AI to render everything from individual buildings to large-scale industrial facilities at appropriate spatial scale. The team also applied a geospatial semantic alignment technique to help the model learn dense structures such as road connectivity and farmland boundaries. In this approach, features extracted by a satellite-imagery-specialized AI serve as a reference for the generative model's training, with the specialized AI used only during the training phase.
After training the common model, the team trained only small sets of additional parameters tailored to cloud removal and SAR-to-optical image conversion. Using the low-rank adaptation (LoRA) method, which keeps most of the core model intact, they reduced the burden of retraining the entire large model for each task.
In cloud removal experiments using publicly available satellite data, the team said the model outperformed existing specialized models on key metrics evaluating brightness, color, and structural similarity. In experiments converting SAR images into optical-style images, it also achieved top scores on some metrics assessing generation quality and fine structural detail, according to the team. Professor Kim said, "We confirmed that the terrain and spatial structures learned during the satellite image generation process are also useful for restoring cloud-obscured imagery and interpreting radar data," adding, "By learning both shooting locations and real-world terrain dimensions, we enabled a single AI to handle multiple tasks such as cloud removal and image transformation."
Dr. Jeonghyeok Do of the KAIST Information and Electronics Research Institute participated as the first author of the study, and according to the team, the results have been accepted at the international AI conference NeurIPS 2026.
