Deep learning solution for the automatic detection of objects from space

In the context of the growing number of satellites that are placed weekly into low Earth orbits (LEO) around the Earth, astronomical images are becoming increasingly contaminated by these objects and other orbital debris, crossing the field of view. Since manually sorting the frames that can be processed for scientific purposes is not a viable solution, a group of researchers from the Astronomical Institute of the Romanian Academy proposes an alternative based on deep learning techniques. The custom model, trained on a consistent set of all-sky images acquired at the Berthelot Observatory (AIRA), achieves 100% precision and 91% sensitivity, meaning that only 9% of objects (satellites or orbital debris) could not be detected, while all detected objects were correctly identified.

From an operational perspective, this combination of high sensitivity and the absence of false alarms is highly advantageous: frames flagged as “contaminated” are sent directly to the SST (Space Surveillance and Tracking) analysis pipeline, while “clean” frames enter the scientific data-processing pipelines without additional manual inspection.

The article published in Astronomy and Computing can be accessed here: https://doi.org/10.1016/j.ascom.2026.101081

Published on: Feb 26, 2026