Pace Balster, Katalyst Space Technologies; Gabrielle Jones, Katalyst Space Technologies; Gavin Hofer, Katalyst Space Technologies; Daniel Newsom, Katalyst Space Technologies; Carolin Frueh, Purdue University;
Keywords: Light Curve, Characterization, Pattern of Life, Classification, Unresolved, Non-Resolved, Shape, Stability, Blob Identification
Abstract:
Optical observations are often limited to non-resolved imagery when observing significantly small or distant objects. These observations are adequate for astrometric measurements used for orbit determination; however, in a densely crowded space, further characterization of space objects beyond center of mass becomes necessary for better orbit prediction, object classification, and pattern of life recognition, all crucial elements to maintain safety of flight of space assets. Collected light curves of space objects from non-resolved imagery presents an opportunity to derive this additional characteristic information. The amount of the reflected light intensity is a superposition of the shape, materials, and attitude of the spacecraft in a given observation geometry and, hence, can offer object characterization insights even in the entirely unresolved case. Complete inversion of the so-called light curves is often not possible in the presence of measurement noise with a small amount of data, even when information such as shape or attitude is known due to the multivariate nature of the problem.
This paper offers an exploration of various methods for object type prediction, focusing on machine learning approaches and a new technique of partial inversion, sometimes called blob identification. The latter involves correlating a short light curve time series with a probability space of interpretations and their associated likelihoods. Both methods are evaluated independently as well as in combination to ascertain the most effective technique for object class determination. The blob identification method clearly distinguishes between characteristic information such as attitude characteristics or parts of reflective surfaces that can be extracted, including the combinatory sets of such information, and information that is inaccessible given the length, cadence, and observation geometry. The machine learning strategy involves a multi-class classification technique, aimed at mapping features within the light-curve to object class probabilities. This paper explores different models, hyperparameters, and feature conditioning in relation to traditional machine learning performance metrics like precision, recall, F1 score, and balanced accuracy. Both the blob identification and machine learning methods are assessed using an analysis architecture that enables rapid evaluation of each approach, using light curves from multiple datasets to reduce bias and to show real-world feasibility. The outputs from these methods allow for the identification of various object types, including actively stabilized and un-stabilized objects, rocket bodies, and debris objects.
The deterministic approach of blob identification employs a sub-parameterization of the complete inversion problem. As such, all dependencies can clearly be outlined as part of the set information, and all results are traceable and reproducible. It also eliminates the need for training sets and the bias they potentially introduce. The seamless transition into the complete inversion problem is possible as more data becomes available, allowing the resolution into finer detail, such as full shape and six-parameter attitude determination. While the deterministic blob identification method is advantageous due to its transparency and lack of reliance on large training sets, machine learning methods offer the capability to uncover hidden relationships within the dataset. A direct comparison of both methods is evaluated as well as a staged approach in which the blob identification method is treated as a pre-processing step to the machine learning method.
A modular analysis framework is developed for the purpose of rapid evaluation of both the partial inversion and machine learning technique. Multiple datasets and edge cases are evaluated to capture variance of real-world observations. The results of the new technique documented in this paper lay the groundwork for future improvements to understand the shape and attitude of space objects given non-resolved imagery.
Distribution A: Approved for public release, distribution is unlimited.Public Affairs release approval AFRL-2023-3763
Date of Conference: September 19-22, 2023
Track: Satellite Characterization