ML-Driven Optimal Design of Multispectral Instruments for the Characterization of Resident Space Objects

The characterization of resident space objects (RSOs) has long been an important aspect of space domain awareness (SDA). Recently, spectral imagery has emerged as a viable sensing modality with the potential to enhance existing RSO-characterization capabilities, especially when used alongside machine-learning (ML) models. Deploying a spaceborne hyperspectral sensor, however, can increase size, weight, power, and cost (SWaP-C). On the other hand, multispectral sensors, which collect significantly fewer bands than hyperspectral sensors, are less complex to develop and deploy. This paper presents an optimization procedure that finds the minimal number of multispectral bands required for training high-performance ML classifiers. The optimization procedure was used to perform a design study for a notional SDA mission responsible for classifying three types of RSOs: active payloads, rocket bodies, and space debris. Simulated hyperspectral signatures of each class were generated and used to initialize the optimal-design procedure. These hyperspectral signatures consisted of 1,574 spectral bands. The procedure was able to design an optimized multispectral system with only five bands. Despite the 99.7% reduction in the number of spectral bands, the overall performance of the ML-driven RSO-classification models only decreased by 1.9%. This result suggests that future SDA missions may be able to replace hyperspectral sensors with bespoke multispectral ones while still meeting requirements related to RSO-characterization capability.