GeoCore-9B AI Model Developed by KAIST to Penetrate Clouds and Darkness in Satellite Imagery
The technology aims to provide reliable Earth observation data for climate monitoring, emergency disaster response, and environmental analysis regardless of weather or time of day.
KAIST Builds Foundation Model for Earth Observation Data
Optical satellites rely on visible light, making them ineffective at night or when thick cloud cover blocks the ground. Synthetic Aperture Radar (SAR) satellites bypass weather and darkness by transmitting radio waves to the surface, but SAR outputs look radically different from standard photographs, complicating manual interpretation. To bridge this gap, the KAIST research team bypassed general-purpose image AI models and built a foundation model from scratch tailored specifically for Earth observation data. The model trains on 10 million pairs of global satellite images and descriptive text, utilizing approximately nine billion parameters.
GeoCore-9B Integrates Geographic Metadata to Minimize Spatial Distortion
The GeoCore-9B architecture integrates traditional image objects with geographic metadata, including latitude and longitude coordinates, Ground Sample Distance (GSD) reflecting actual ground scale, and structural boundaries such as road networks and agricultural fields. This precise combination minimizes spatial distortion relative to actual terrain. In performance evaluations covering cloud removal and SAR-to-optical image translation, the model outperformed existing specialized architectures. The research team also incorporated Low-Rank Adaptation (LoRA) efficiency techniques, allowing the system to adjust select parameters without requiring full retraining of the entire nine-billion-parameter network.
KAIST, 위성영상 파운데이션 AI '지오코어-9B' (머니투데이, 10월 20일)1. KAIST 김문철 교수 연구팀: 위성영상 생성·복원·변환용 AI '지오코어-9B(GeoCore-9B)' 개발2. 매개변수 약 90억 개: 위성영상-설명 문장 1000만 쌍 학습, 위도·경도·지상표본거리(GSD) 반영3. 기능: 구름 가린 영상… pic.twitter.com/qxTCwQro8p
Applications in Disaster Response and National Defense
Kim noted that eliminating blind spots during severe weather and nighttime operations will support rapid disaster response, environmental tracking, and national security surveillance. The research, featuring Dr. Jeong-Hyeok Do of the KAIST Information Electronics Research Institute as the first author, has been officially accepted for presentation at the NeurIPS 2026 artificial intelligence conference.
Keep reading