Spotting a tumbling satellite from the light it reflects: 88% detection

A model trained on the light curves recorded by telescopes distinguishes a rotating satellite from a tumbled satellite, and signals 88% of abnormal behaviour to operators.

An international consortium — University of Strathclyde, Alan Turing Institute, University of Arizona, MIT, University of Waterloo, with the manufacturers GMV, Columbiad, LMO and Zendir — presented a model in early September 2026 that predicts anomalies and the motion of objects in orbit by exploiting only one piece of data: the way they reflect light.

These "light curves," taken from ground-based telescopes, contain complex information about a satellite's attitude and behavior. The model was trained on large volumes of real readings and then refined on selected simulation data. Fed with recent or real-time curves, it flags atypical patterns for a human expert to examine. In tests, it identifies abnormal curves in 88% of cases and can distinguish a rotation from a tumble—a distinction that directly affects in-orbit service and life extension operations.

The work is part of the AI4 Space Safety and Sustainability project, led by Professor Massimiliano Vasile and funded by the British Space Agency as part of a consortium involving Australia, Canada, New Zealand, the United Kingdom and the United States. The results have been published in the journal Expert Systems. The next step is to add radar, hyperspectral and orbital data.