Kedar Naik, BAE Systems, Space & Mission Systems; Andrew Wernersbach, BAE Systems, Space & Mission Systems; Alexandra Robinson, BAE Systems, Space & Mission Systems; Michelle Nilson, BAE Systems, Space & Mission Systems; Gary Wiemokly, BAE Systems, Space & Mission Systems; Matthew Tooth, BAE Systems, Space & Mission Systems; Raymond Wright, BAE Systems, Space & Mission Systems
Keywords: Machine learning, design optimization, material characterization, hyperspectral, multispectral, SDA, RSO classification
Abstract:
Over the past couple decades, several studies have established the viability and the benefit of hyperspectral imagery in space-domain awareness (SDA) [1]. More recently, other studies [2, 3] have demonstrated the ability of machine learning (ML) models to characterize the material composition of resident space objects (RSOs). One problem that has received less attention in the literature, however, is the high cost associated with designing and deploying a hyperspectral imager. The proposed paper seeks to address this problem while simultaneously building upon previous work related to ML-based spectral sensing for SDA missions. Specifically, this paper will present an optimal-design procedure capable of replacing a hyperspectral sensor with a cheaper, multispectral one, without sacrificing the accuracy of ML models trained to classify common types of RSOs. For the present study, three classes of RSOs are considered, namely: active payloads, abandoned rocket bodies, and space debris. The design procedure allows the user to supply an arbitrary number of mission-relevant RSO classes and returns (a) the minimal set of spectral bands required to classify those RSOs and (b) the ML models capable of accurately making those classifications. This minimal set of band centers and bandwidths comprises the most critical information required to design the optical filters required to build a complete multispectral system.
The design procedure to be presented in this paper relies on four key algorithmic components, viz., (1) a astrodynamics-informed spectral-signature data generator, (2) an explainable, pixelwise ML-training suite, (3) a spectral-downsampling tool, and (4) a derivative-free optimization routine. All four of these components are used iteratively. Before describing the mechanics of the design procedure itself, it is worth elaborating on the four key algorithmic components mentioned above. The first is a software package called COAST that allows the user to generate an arbitrary number of simulated spectral signatures for each mission-relevant class of RSO. COAST employs an orbital-mechanics simulation in conjunction with a radiometric model and a materials database to generate thousands of synthetic spectra for each RSO class of interest. That large collection of synthetic spectra can then be used to train classification models using ML. The second component, a software suite called Classification Of Pixelwise Spectra (COPS), is capable of automatically training optimal ML-classification models on pixelwise spectral data. Specifically, COPS produces an ensemble model of gradient-boosted decision trees for each RSOs class. These ensemble models are considered optimal inasmuch as their hyperparameters are tuned using a Bayesian-optimization routine. An important benefit of using gradient-boosted decision trees is the ability to quickly extract feature-importance values from the model. (N.B. Here, the features are spectral bands, so the feature-importance values simply quantify the relative importance of each band to the model’s overall ability to make an accurate classification.) The third component is a spectral-downsampling tool that allows a high-resolution spectral signature (e.g., one captured by a hyperspectral imager) to be interpolated to a lower-resolution spectral grid (e.g., a specified set of bandpass filters) in a physically meaningful way. This is done by first constructing a unique Gaussian function to approximate each filter’s spectral profile and then performing a discrete convolution over the higher-resolution spectrum. The fourth and last key component is a derivative-free optimization algorithm. For the present study, a Bayesian optimization method will be used along with constraints that prevent non-physical sets of filters from being suggested.
Having introduced each of the four key algorithmic components, the overall design procedure can be described in a relatively straightforward manner. The process begins with the user selecting a set RSOs classes of interest. Using representative geometries for each, thousands of synthetic spectra are generated using the COAST software. These spectra are then fed to the COPS suite to train optimized binary classifiers for each RSO class. The accuracy of each trained ML model is computed on a held-out test set and the feature-importance values are extracted. Wavelengths with high feature importance are retained, whereas those with low feature importance are discarded. The optimal bandwidths corresponding to each retained wavelength are solved for using Bayesian optimization. This process of model-training, feature exclusion, and bandwidth optimization is iterated – i.e., the number of included wavelengths continues to be reduced – until the aggregate detection accuracy of the ML models is just shy of no longer satisfying mission requirements. The end results are (1) the minimal set of bands required for designing a multispectral RSO-classification system and (2) the ML models required for doing that characterization.
In addition to explaining the technical details of the key algorithmic components as well as those of the design procedure itself, this paper will present a step-by-step demonstration of how an initial list of RSO classes is processed to yield an optimized set of optical filters and ML models. The authors believe that the proposed paper will provide the SDA community with a new process for designing multispectral imagers, which – with the assistance of ML models – are capable of meeting mission requirements in a cost-effective way.
References:
[1] Chaudhary, Anil B., Birkemeier, C., Gregory, S. A., Payne, T.E., and Brown, J. A. “Unmixing the materials and mechanics contributions in non-resolved object signatures,” Advanced Maui Optical and Space Surveillance Technologies Conference, Maui, HI, 2008.
[2] Vasile, M., Walker, L., Dunphy, D., Zabalza, J., Murray, P., Marshall, S., and Savitski, V. “Intelligent characterisation of space objects with hyperspectral imaging,” Acta Astronautica, 203, 510-534, 2022.
[3] Vasile, M., Walker, L., Campbell, A., Marto, S., Murray, P., Marshall, S., and Savitski, V., “Space object identification and classification from hyperspectral material analysis, Scientific Reports, 14, 1570, 2024.
Date of Conference: September 17-20, 2024
Track: SDA Systems & Instrumentation