Kedar Naik, BAE Systems; Andrew Wernersbach, BAE Systems; Nathan Carlson, BAE Systems; Thomas Drouillard, BAE Systems; Raymond Wright, BAE Systems
https://doi.org/10.64861/DDJD2613
Keywords: Multispectral Instrument Design, Satellite Characterization, Spectral Unmixing, Design Optimization, Machine Learning, Explainable AI
Abstract:
As the number of resident space objects (RSOs) in the GEO belt increases, so does the need for long-range RSO characterization that is both fast and informative. Speed is required due to the expansive volume to be monitored; informative characterization implies the ability to infer an object’s origin, intent, and status. Estimating the material composition of an RSO can help provide insight into all three of these characteristics. The presence of certain alloys or polymers can point to national origin; the combination of certain materials can imply mission intent; and the relative abundance of each material at a given point in time can be indicative of status. Remotely sensing material composition – especially for RSOs in GEO – however, has remained a longstanding challenge for the Space Domain Awareness (SDA) community. As a consequence of the vast distances involved, capturing resolved imagery of the GEO belt is generally infeasible with ground-based sensors, which instead rely primarily on unresolved photometric observations. Even within the context of space-based SDA, the radiometry is challenging. Indeed, for a standard LEO-to-GEO wide-area search mission, maximizing the signal-to-noise ratio (SNR) for distant, dim targets within limited dwell times requires deliberately using a low sampling ratio (Q < 1). In such cases, an RSO signal is typically concentrated within a single pixel, thereby eliminating all geometric information and precluding the use of traditional profile-based classification and identification algorithms. One well-established approach for extracting more information from a single pixel is hyperspectral sensing, wherein a spectral signature is captured across multiple wavelength bands within each spatial pixel. Deploying a spaceborne hyperspectral sensor, however, can increase size, weight, power, and cost. On the other hand, multispectral sensors, which collect significantly fewer bands than hyperspectral sensors, are less complex to develop and deploy. Moreover, in recent years, artificial intelligence and machine-learning (AI/ML) models, with their ability to discover and exploit subtle patterns in noisy data, have been successfully used to process spectral data. So, in attempting to replace hyperspectral sensing with multispectral sensing, a pair of design questions naturally arises: (1) assuming ML models can be used for data processing, what is the minimum number of spectral filters required for estimating material composition and (2) what are the precise wavelengths and bandwidths of those filters? To that end, this paper presents an optimization procedure that finds the minimum number of multispectral bands required for training high-performance deep-learning models for single-pixel material-composition estimation. Building upon a 2024 AMOS publication by the present authors [1], which introduced the MOB BOSS (Mission-Oriented Bayesian Band Optimization for Spectral Systems) framework for AI-driven instrument design and broad RSO classification, this paper introduces an expanded and upgraded version of MOB BOSS focused on sub-pixel material-abundance estimation. At a conceptual level, MOB BOSS is an optimization loop that returns the minimal set of multispectral filters – as well as the AI/ML models trained to process the corresponding multispectral signatures – required to satisfy a specific mission objective (RSO classification, material-composition estimation, etc.). It functions by training an initial machine learning (ML) model on high-dimensional hyperspectral data, employing explainable AI (XAI) techniques to identify the most influential spectral ranges, extracting from those ranges a reduced set of bands, and then using a Bayesian-optimization routine to iteratively retrain the ML model while refining each band’s center wavelength and bandwidth. This process of band-set extraction, model training, and band refinement continues until the number of multispectral filters converges. Our 2024 paper focused on RSO classification; to enable sub-pixel material-composition extraction, we integrate two recent advancements: a one-dimensional convolutional neural network (CNN) architecture for high-accuracy spectral unmixing [2] and a demonstrated capability to distinguish complex material sets from single-pixel signatures [3]. The CNN architecture employed in the present study makes use of a multiple-regression head and a bespoke loss function to predict abundance fractions for each entry in a library of common aerospace materials (e.g., Kapton, solar cells, and various paints). To extract influential spectral ranges from the CNN model, a new XAI method is required. We introduce Vector-Target Integrated Gradients (VTIG) to assign feature-importance values to each spectral band. While traditional Integrated Gradients [4] provide attributions for scalar outputs, VTIG explains the entire set of material-abundance predictions simultaneously. By looping the IG method over each node in the regression head and calculating the L2 norm across the output dimension, VTIG identifies which spectral bands move the entire set of material-abundance predictions the most. This allows the MOB BOSS framework to underweight spectral bands that do not contribute to unmixing accuracy. As with the initial MOB BOSS publication, hyperspectral training data is generated using the COAST physics-based simulation environment, which allows for the generation of high-fidelity, noisy spectral mixtures from a variety of solar-incidence angles and observer positions. In summary, this paper will demonstrate that the MOB BOSS framework – augmented with CNN regression models and VTIG – can optimally design the lowest-cost multispectral sensor required to perform RSO material-composition estimation, returning both the optimal filter specifications and a mission-ready, trained neural network tailored to process data captured by those specific filters.
Date of Conference: September 15-18, 2026
Track: Satellite Characterization