Clarice Reid, Slingshot Aerospace; David Witman, Slingshot Aerospace; Timothy Olson, Slingshot Aerospace; Carl Steinhauser, Slingshot Aerospace; Aaron Pung, Slingshot Aerospace; Brian Williams, Slingshot Aerospace; Dylan Kesler, Slingshot Aerospace; Belinda Marchand, Slingshot Aerospace
Keywords: Light-curve Inversion, Fingerprinting, Attitude Change, Photometric Analysis, Machine Learning
Abstract:
The observational data generated by terrestrial sensors is rapidly expanding alongside the proliferation of satellites in low Earth orbit (LEO). Understanding and extracting information from these sensor systems is paramount to creating an accurate representation of the orbital arena. Challenges include object identification and geometric change detection. Accurate object identification using fingerprinting enables downstream algorithms to better correlate objects, evaluate relationships among objects, and enable categorization of object geometries. Change detection algorithms are intended to identify shifts in attitude or geometry (shape) that may foretell impending maneuvers, separation events, or different orientation modes. Current workflows often include manual examinations of photometric data by orbital analysts, but automated methods must be developed to scale as the number of objects in LEO grows. In this paper we present an overview of a method for addressing both object fingerprinting and attitude change detection and demonstrate it at scale.
Ground based Electro-Optical (EO) imagery of LEO space objects has advanced to the point that millions of observations can be produced over the course of a single day. Along with these observations, many derived features, like photometric magnitude and sun-satellite-sensor angles, can be extracted to augment and contextualize the information contained in any one observation. Distributed sensor networks provide many observation opportunities that span an angular feature space, such that extensive reflectance patterns can be captured on a single object. Similarly, developed models, state, and contextual information can be combined to predict photometric returns based on given relative positions, object attitudes and geometries, and reflectance angles due to the sun, moon and earth.
We combine an extensive observation dataset with a novel machine learning (ML) model and contextual information to develop a method that differentiates between objects with distinct geometries. Since true object geometries are often not shared publicly, we develop a generic representation using object bus types and constellation membership to define our ML classification problem. Previous work on photometric fingerprinting showed promising performance of a twin neural network to discriminate between Starlink bus classes. We extend this work with our new method to consider a greatly expanded set of classes and to capture more information on sparser data. These models are evaluated across a representative sample of LEO objects that span the angular observation space, capture a wide variety of bus types, and include a multiple satellite operators and manufacturers.
Our fingerprinting model extends to the problem of attitude change detection using the learned representations of the data over time. Robust object representations built over multiple viewing angles provide inputs that allow us to evaluate object behavior (e.g. attitude). Applying change point detection methods on these assignments allows us to back out stability or attitude changes. Truth data is difficult to recover in this field, as attitude is rarely provided by operators, and spin rates of debris or other objects is often unknown. We address this both by observing changes in embedded representations, and also by simulating representative light curves with different geometries and attitude changes at different illumination conditions. The first method allows for a data-driven look at the changes in object representations over time, and the second allows for a controlled approach to change detection and labelling.
The ultimate goal of this paper is to expand on prior work in photometric fingerprinting to include more object geometries, to capture changes in attitude, and to increase model flexibility and reduce datapoints needed for useful fingerprints. We share a background on the derived features that were considered to build the proposed model. Details on the model are provided along with discussion of its strengths and weaknesses. The proposed model explores representations that apply to both object identification (fingerprinting) and attitude change detection. Finally, performance results on the object identification task are shared along with preliminary results on the ability of these models to detect attitude changes in real and simulated data.
Date of Conference: September 16-19, 2025
Track: Satellite Characterization