Alexander Agathanggelou, Defence Science and Technology Laboratory (DSTL); Ryan Houghton, Defence Science and Technology Laboratory (DSTL); Joshua Collyer, Defence Science and Technology Laboratory (DSTL); Joshua Davis, Defence Science and Technology Laboratory (DSTL); Nicholas Pallecaros, Defence Science and Technology Laboratory (DSTL)
Keywords: Space Domain Awareness, Machine Learning, Satellite Characterisation
Abstract:
The space domain is becoming ever more congested. The number of satellite launches is growing rapidly; commercialisation of Space Surveillance and Tracking (SST) services and the number of space-capable nations continues to rise; and on-orbit collisions between active satellites and other objects is an ever present and growing risk. In tandem, the amount of observational and technical data available on Earth satellites is growing, delivering a more complex but potentially more veracious view of the situation. These two factors increasing domain congestion and an overload of data, presents a significant challenge to contemporary orbital analysts attempting to maintain the standards of space domain awareness and services that were more achievable in past decades. However, it is shown that Machine Learning (ML) techniques can potentially offer assistance to analysts in managing these burdens: ML can leverage high volumes of data to provide timesaving, wider-reaching, or novel capabilities, and release human effort to focus on complex or higher-priority issues.
This paper demonstrates the feasibility and effectiveness of machine learning models for satellite characterisation. Dstl have developed models that successfully predict the operational status of satellites by measuring fluctuations in reflected sunlight while they orbit overhead at night, or similarly fluctuations in reflected radar pulses (with an accuracy of around 86% for LEO, and 92% for GEO). Notably, the accuracy of the ML assessment only dropped by a few percent when only presented with a single track (observed overhead pass) rather than the entire dataset, demonstrating the utility of even a single light curve collection when using these characterisation techniques. We also developed machine learning models to identify the bus type of a satellite from the same data (with an accuracy of >98%).
The best results in both cases were achieved when the model was provided with additional context, such as the launch date of the satellite. We found classical ML models (e.g. Random Forests) performed just as well as Deep Learning approaches: our LSTM (long short-term memory) models struggled to retain their high performance when generalising to non-iid (independent identically distributed), out-of-distribution samples. The full range of characterisation tasks tackled by analysts is significantly broader than those investigated here, and considerable scope remains for further research.
Data availability was a critical and limiting factor throughout this work. Based on the sources available, we found novel Wide Field Lens Arrays telescopes to be well suited to characterising LEO satellites: the large field of view and rapid exposure rates provided the highest quality measurements and the best model performance on closer and brighter satellites, but struggled with the fainter GEO reflections. However, data from such sensors are not currently available as a commercial service. Conversely we found that existing commercial world-wide networks of traditional telescopes were ideal for characterising GEO satellites: placing facilities across the globe allows the entire GEO belt to be observed and traditional Newtonian telescopes easily record faint GEO reflections.
Open-source labels and information pertaining to current satellite operational statuses were not always available to high enough accuracy, limiting model performance. In many cases, different sources of satellite information (open-source and commercial) did not always agree, particularly with regards to operational status. This limited the use of standard supervised ML models, which cannot easily exceed the accuracy of the labels in their training data. Furthermore, these sources only provide data associated with the current state of the satellite; there is no publically available archive of satellite information, which hampers labelling (and use) of historic observations for model training purposes.
The application of machine learning to satellite characterisation is an emerging nascent topic. While we have demonstrated capability that could maintain the standards of products and services we have become accustomed to, considerable scope remains to further develop and expand this work. As space analysts become increasingly overwhelmed, compromises in what products they offer are already being considered: such concessions might not even be necessary if we can rapidly augment analysts with wider reaching, machine learning powered capabilities.
Date of Conference: September 27-20, 2022
Track: Machine Learning for SSA Applications