A satellite's solar panel or robotic arm can disappear behind its body as the viewing angle changes. Researchers at the Alan Turing Institute and the University of Birmingham have developed an AI method that uses earlier radar views to help interpret later ones. In a simulated study, this temporal memory improved the model's ability to label the spacecraft's components.[1][2]
The institute announced the work on 9 October for the European Radar Conference. Its technical paper uses images generated by the GEIST radar simulator from one modified Clementine satellite design, with a robotic arm replacing a solar panel. This is a study of simulated space-to-space radar observations, with satellite inspection and servicing as prospective applications.[2]
A map of parts, pixel by pixel
Semantic segmentation assigns image pixels to categories. For satellite radar, those categories can describe the body, panels, antennas, propulsion equipment or other structures. A component map provides more detail than an object's position in orbit and could help an operator interpret its shape during an approach.[1][2]
The approach uses high-frequency inverse synthetic aperture radar, or ISAR. Relative motion and radar processing produce views of the target's structure. The model then has to connect those views with meaningful components, despite changes in scattering and which surfaces are visible.[1][2]
Imagine an antenna clearly separated from the body in one view, then partially obscured in the next. Treating each image independently discards the earlier clue. The study adds a recurrent memory component to a U-Net segmentation network, allowing information to pass along the sequence.[2]
What the experiments measured
Results on the simulated satellite
| Configuration | Mean IoU |
|---|---|
| Seven groups, per-frame baseline | 0.581 |
| Seven groups, with class weighting | 0.642 |
| Seven groups, weighting plus temporal memory | 0.68 |
| Thirty classes, per-frame model | 0.44 |
The progression helps explain where the improvement came from. Class weighting increased the importance of underrepresented parts during training. Temporal context then added information across views. The headline 0.68 result applies to seven broader component groups, while the finer 30-class task was more difficult.[2]
Small parts occupy fewer pixels and offer less evidence for classification. The paper identifies them as a persistent challenge, alongside the limited range of spacecraft designs in the dataset. A result on one known geometry does not yet establish recognition of unfamiliar or damaged satellites.[2]
From component maps to orbital operations
The institute describes possible uses in monitoring, rendezvous, maintenance and debris removal. Those activities need a detailed account of where structures are and how they move. A component map could become one input alongside orbit estimates and other observations; operational navigation requires additional validation and controls.[1]
The scientific interest is the use of changing views as cumulative evidence. Rotation can hide one feature while exposing another, so a sequence may contain more useful structure than its individual images suggest. The study quantifies that benefit in a controlled simulation and identifies the small components and unfamiliar geometries that a broader model would need to handle.[1][2]
Sources & further reading
- New AI model will boost UK's ability to safeguard vital satellites The Alan Turing Institute · 9 Oct 2026 · Accessed 27 Oct 2026
- Semantic Learning of Satellite Morphology using in-orbit Inverse Synthetic Aperture Radar Alan Turing Institute / European Radar Conference 2026 · Accessed 27 Oct 2026