Alexander Macmanus, University of Warwick & Defence Science Technology Laboratory (Dstl); Don Pollacco, University of Warwick; Paul Chote, University of Warwick; Calum Meredith, Defence Science Technology Laboratory (Dstl)
Keywords: SSA, SDA, characterisation, LEO, optical, infrared, machine learning
Abstract:
There has been a recent increase in the growth of the number of resident space objects (RSOs) and as a result, the importance of Space Domain Awareness (SDA) has never been greater. There are around 45,000 objects in the public catalogue, not including objects too small to be tracked and any satellites deliberately omitted from the catalogue, such as some military satellites. Estimates of untracked objects vary, but estimates have run as high as 100 trillion. This quantity of data threatens to overwhelm current analyst capacity, and so the focus of this research is on developing methods and tools in a subset of SDA concerned with the characterisation of existing tracked objects, using machine learning to aid analysts in bulk data processing. This characterisation can involve the calculation of simple metrics such as spin axis, spin rate, orbital height and stability, or estimation of albedo, but it can also involve combining multiple metrics into a more comprehensive analysis, trying to determine characteristics like bus type, mission type, geometry, or payload. It can also involve the identification of particular equipment on a satellite, such as telecom transmitters and receivers, radar dishes, and most frequently solar panels. This research aims only to target Low-Earth Orbit (LEO) satellites, with altitudes no greater than 2000km. This presents a markedly different challenge than at Geostationary Orbit (GEO) – not only is there a tracking requirement, but the observation time for a given object is typically limited to an observing window of less than 10 minutes as the object passes from horizon to horizon. This practically limits the light curves (the object’s brightness and its change over time) to at most 10-minute segments, with no guarantee that the object will be visible again in the same night.
This paper presents a brand new dataset of LEO light curves collected at the Roque de los Muchachos observatory on La Palma in the Canary Islands, using the CLASP (Challenge Lead Applied Systems Program) telescope owned and operated by the University of Warwick. CLASP comprises two co-mounted optical tubes, with an optical CMOS sensor on one and a short-wave IR sensor on the other. Between 40 and 80 light curves were collected per night of observations depending on the weather and the time of year, both of which govern the available observing window. These light curves are generated automatically from the raw images using in-house tooling developed to identify the satellites in the frame and generate the photometry – this was done principally to keep the data throughput high and the long-term storage requirements low. The feasibility and drawbacks of this approach are discussed in the paper. This dataset consists of 1227 single-band optical light curves covering 901 different satellites (collected 2024) and an 1886 contemporaneous dual-band optical/short-wave infrared light curves covering 940 different satellites (collected 2025), with more being added continuously. The contemporaneous collection of data in both optical and short-wave infrared establishes a new layer of analysis involving the fusion of the two – some features of one light curve are either not present in the other, or shifted by some fixed time, allowing for a greater understanding of the satellite hardware and operation to be determined than is possible from single band data.
Analysis of the dataset was also undertaken using traditional machine learning techniques – in particular the use of decision trees on light curve features for the purposes of classification, using Seradata labels as ground truth labels for the data. Additional analysis is provided for an initial look into the relative performance gained using dual-band data over exclusively EO or SWIR data.
In summary this paper presents a new method for scalable, automated light curve collection & processing for LEO RSOs in SWIR, and presents an analysis of the accuracy of selected machine learning algorithms as applied to the currently collected light curves in the database. This will greatly reduce operator burden by allowing characterisation features to be automatically extracted at scale.
Dstl Crown Copyright © 2025
Date of Conference: September 16-19, 2025
Track: Satellite Characterization