Classifying LEO satellite platforms with boosted decision trees

Abstract

Abstract As the cost of reaching LEO (low Earth orbit) has diminished we expect, over the next decade, an almost exponential increase in the numbers of LEO spacecraft from established and potentially new agents. Remote characterization of these and the increasing numbers of decommissioned/debris satellites are thus becoming more important, along with identifying unannounced changes in megaconstellations. In this paper we examine the light curves of known LEO platforms with a boosted tree algorithm in order to determine whether spacecraft properties were discernible. A priori we expected little correlation as we expected the large variations in sight-line geometries would mask signs from the spacecraft. Using large numbers of light curves from the MMT-9 data base, we find that this is not the case and most platforms are statistically identifiable in most sight-lines and tentatively associate this correlation with the differences and similarities between downward facing instruments. Pairs of satellite platforms can be distinguished 86.13 per cent ($N=15\, 600$) of the time using this method. Evolutionary changes to the Starlink satellite platform are also distinguished.