Satellite images of the Earth are not always easy to use. Clouds can block the view of the ground, and images captured with certain imaging methods look different from ordinary photographs, which makes them difficult to interpret. A KAIST research team has developed an artificial intelligence (AI) model that learns the appearance of diverse terrain and cities. The model can restore cloud-covered images and convert radar images into images that resemble ordinary photographs. It applies what it learned while generating satellite images to a range of image-processing tasks, so it is expected to expand the use of satellite data in disaster response and environmental monitoring.

KAIST (President Choongsik Bae) announced on October 20 that a research team led by Professor Munchurl Kim of the School of Electrical Engineering has developed GeoCore-9B, an AI model that can be used to generate, restore, and convert satellite images. During training, the model’s core image-generating component learns image features by adjusting about 9 billion internal values, known as parameters. GeoCore-9B is pretrained on a large volume of satellite data and can then be adapted to a variety of tasks. For this reason, it is classified as a ‘foundation model,’ a type of AI model that serves as the basis for a wide range of AI technologies.

Satellite images are essential for monitoring changes in forests and cities and for responding to disasters such as floods and wildfires. However, optical images, which record sunlight reflected from the Earth’s surface, cannot clearly capture the ground when clouds are thick or at night. Synthetic aperture radar (SAR) observes the ground by transmitting radio waves and measuring the signals reflected back. It can acquire images of the Earth’s surface at night and in cloudy weather. However, SAR images differ from ordinary photographs in brightness and patterns, which makes them difficult to interpret.

To create a common basis for processing these different types of satellite data, the research team trained the core image-generating model from scratch on Earth observation data. Instead of adapting the core component of a model built to generate ordinary photographs, the team trained a new model on overhead satellite views of terrain, forests, ports, and cities.

The model was trained on 10 million pairs of satellite images from around the world and sentences describing each image. For example, an image of a port was paired with a description stating that ships and warehouses are located there. An image of a rural area was paired with a description stating that a road runs between fields. By learning from this large collection of described images, the AI learned the features of diverse terrain and urban areas.

A key feature of this study is that the AI learned not only what an image contains, but also which location on Earth the image shows and how large an area it covers. In satellite images, objects with similar shapes may be small buildings or large industrial facilities, depending on their actual size. It is therefore important to take into account both where an image was taken and its spatial scale.

The team trained the model using latitude, longitude, and ground sample distance (GSD) together with the descriptive sentences. Latitude and longitude indicate a position on Earth. GSD indicates the size of the ground area represented by a single pixel, the smallest unit of an image. For example, at a GSD of 1 meter, one pixel represents an area of about 1 meter by 1 meter on the ground. Using this information, the AI learned to generate images that match a given location and spatial scale.

The team also applied a training technique called ‘Geospatial Semantic Alignment,’ which helps the model represent the structure of terrain and cities more accurately. In this technique, the generative model learns by referring to features extracted by an AI specialized in satellite imagery. These features serve as a training reference that helps the model learn structural elements of an image, such as how roads connect and where farmland boundaries lie. The specialized AI is used only during training, so it does not add computation when the model generates images.

The researchers then fine-tuned GeoCore-9B separately for cloud removal and for radar image conversion. They used LoRA (Low-Rank Adaptation), a method that keeps the core model fixed and trains only a small number of additional parameters. This approach reduces the burden of retraining an entire large AI model from scratch for each task.

In evaluations using publicly available satellite datasets, GeoCore-9B outperformed existing task-specific models at restoring cloud-covered images. It scored higher on the key metrics that measure brightness, color, and structural similarity to the reference images. In converting radar images into optical images, it also achieved the best results on some of the metrics that assess the quality and fine detail of the generated images.

These results demonstrate that knowledge gained from generating satellite images can also be applied to restoring and converting actual observation images. The model is expected to make greater use of data whose value has been limited by cloud cover or imaging methods. It is also expected to support the assessment of conditions in disaster areas and the monitoring of changes in forests, farmland, and cities. In addition, adapting a single common model to multiple tasks can reduce the cost and time required to develop a separate AI for each task.

The team plans to validate the model’s performance across a wider range of sensors and regions. To apply the model to real-world disaster response, the team also plans to expand its research to accurately predicting changes on the ground, such as floods, building damage, wildfires, and urban expansion. The team will also explore defense applications such as surveillance and reconnaissance.

“By training the AI on where each satellite image was taken and the actual size of the terrain it shows, we made it possible to use a single model for multiple tasks such as cloud removal and image conversion,” said Professor Kim. He added that he expects the work to make satellite data limited by weather or imaging methods more widely usable, which will help with assessing disaster situations and monitoring changes in forests and cities.

Dr. Jeonghyeok Do of KAIST’s Information & Electronics Research Institute led the study as first author, and Professor Munchurl Kim of the School of Electrical Engineering participated as corresponding author. The findings have been accepted for presentation at NeurIPS 2026, a major international conference in artificial intelligence. The paper was posted on the preprint server arXiv on August 3, 2026.
※ Paper title: GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation. DOI: https://doi.org/10.48550/arXiv.2608.01896
※ Authors: Jeonghyeok Do, Munchurl Kim

This research was carried out using high-performance computing resources provided through the ‘Advanced GPU Utilization Support Program’ of the Ministry of Science and ICT and the National IT Industry Promotion Agency. GPUs are semiconductor chips that process a large number of calculations at the same time, and they are used to train large-scale AI models.