Miguel Velez-Reyes, The University of Texas at El Paso; Aryzbe Najera, University of Texas-El Paso; Leah Porras, The University of Texas at El Paso; Dan DeBlasio, University of Texas at El Paso; Hector Erives, The University of Texas at El Paso
Keywords: remote sensing, unresolved resident space objects, hyperspectral, spectro-temporal signatures, signature analysis, simulation, machine learning, training
Abstract:
Abstract
Advancing SDA to provide tactical, predictive, and intelligence information on resident space objects (RSO) will rely on successfully extracting information from ground-based remote sensing assets. Radar and optical remote sensing systems are the primary assets for ground-based observations [1]. Radar is primarily used for targets in LEO while optical assets are used for observations at higher orbits.
The unresolved RSO (URSO) problem arises because optical remote sensing assets cannot spatially resolve RSO that are far away (e.g. GEO) or that are small (e.g. nanosatellites). Furthermore, the low cost of small aperture COTS telescopes is motivating their use for observation of LEO satellites with high spatial diversity even though they cannot spatially resolve them [2, 3, 4].
Standard astronomical measurements such as photometry, spectroscopy, and polarimetry have been used since the beginning of the space age to characterize a satellite’s optical signature [5]. Signature variability is due to the superposition of object shape, attitude, motion, and material composition under a specific viewing and illumination geometry [6]. Hyperspectral and polarization remote sensing systems [7, 8] and the geographical diversity provided by assets such as the USAFA Falcon Telescope Network [9] or OWL-Net [4] are bringing enormous amounts of data that can provide further insights into aspects of URSO that were previously inaccessible. Accurate signature interpretation may allow us to perceive, predict, comprehend, and react appropriately to changing situations in the space domain.
Hyperspectral sensors allow extraction of URSO information from the spectro-temporal (ST) variability of measured spectra providing a quantitative approach for characterization. Even though the URSO cannot be spatially resolved, many of its properties can be extracted from the spectro-temporal information.
Recent studies show promise for leveraging machine-learning (ML) techniques to infer satellite properties such as operational status [10], orbital regime [11], and shape [12] from light curves. These ML models exhibit strong performance when trained and tested on simulated data from simple simulation models. However, ML models trained on this simulated data have significant reduction (up to 33% in accuracy) when applied to real observations [11, 12]. High-fidelity physics-based models for spectro-temporal signature generation can yield significant improvements in ML model performance, and lead to predictable false-alarm rates in certain observer-illuminator geometries [13, 14]. Simulated datasets are key for training of ML models for SDA as availability of real data is rather limited.
The Digital Imaging and Remote Sensing Image Generation (DIRSIGTM) model is a tool that produces synthetic imagery using a suite of physics-based radiation propagation modules. DIRSIGTM can simulate images taken from different sensors with variation in collection geometry, spectral response, solar elevation and angle, atmospheric models, target, and background. DIRSIGTM has been applied to RSO and URSO simulation scenarios in [15] and has been shown to be a valuable tool for generating synthetic data for training of deep convolutional networks for satellite or aerial image analysis as shown in [16].
In order to interpret ground-based spectro-temporal signatures to solve information extraction problems in RSO remote sensing, we need to understand their dependency on operational conditions including the geographic location and angular resolution of the telescope, observation geometry, satellite attitude and orbital characteristics, materials in the surface of the RSO, atmospheric conditions among other variables. We can develop this knowledge by using physics-based simulation models.
In this paper, we will present a simulation-based study of the spectral variability of ground-based RSO spectro-temporal signatures under imaging collection conditions of interest. The experiment is helping us develop an understanding of the dependency of spectro-temporal signatures on these conditions and simulated data sets that can complement ground-based observations for unsupervised training of a convolutional variational autoencoder to produce an accurate low-dimensional latent representation of the RSO spectro-temporal signature data. This representation can be used then to train prediction networks for key aspects of interest from the RSO (e.g. spin rates, anomalies, and RSO classification).
The paper will also describe the simulation workflow, needed ancillary data, and resources needed to implement the simulation model in the DIRSIGTM environment.
References
[1] S. Eves, Space Traffic Control, vol. 251, AAIA, 2017.
[2] N. Estekk, et al., “Daylight imaging of LEO satellites using COTS hardware,” AMOS Conference, 2019.
[3] F. Gasdia, et al., “Multi-Site Simultaneous Time-Resolved Photometry with a Low Cost Electro-Optics System,” Sensors, vol. 17, p. 1239, May 2017.
[4] J.-H. Park, et al., “OWL-Net: A global network of robotic telescopes for satellite observation,” Advances in Space Research, vol. 62, p. 152–163, July 2018.
[5] R. Tousey, “Optical Problems of the Satellite,” Journal of the Optical Society of America, vol. 47, p. 261, 4 1957.
[6] C. Fruh and T. Schildknecht, “Analysis of observed and simulated light curves of space debris,” Proceedings of the International Astronautical Congress, 2010.
[7] M. Pirozzoli, et al., “Calibration and sensitivity analysis of a basic polarimeter for manmade satellite observations,” AMOS Conference, 2020.
[8] A. Speicher, et al., “Results from an experiment that collected visible-light polarization data using unresolved imagery for classification of geosynchronous satellites,” Airborne Intelligence, Surveillance, Reconnaissance (ISR) Systems and Applications XII, 2015.
[9] F. K. Chun, et al., “A new global array of optical telescopes: The falcon telescope network,” Publications of the Astronomical Society of the Pacific, vol. 130, p. 095003, 2018.
[10] J. R. Chapman, et al., “Wavelet Covariance Analysis for Light Curve Slew Maneuver Detection,” AIAA Scitech 2020 Forum, 2020.
[11] R. Linares and R. Furfaro, “Space Object classification using deep Convolutional Neural Networks,” 19th International Conference on Information Fusion (FUSION), 2016.
[12] R. Furfaro, et al., “Shape Identification of Space Objects via Light Curve Inversion Using Deep Learning Models,” AMOS Conference, 2019.
[13] G. Badura, et al., “Convolutional Neural Networks for Inference of Space Object Attitude Status,” AMOS Conference, 2020.
[14] D. Bennett, et al., “SSA Modeling and Simulation with DIRSIG,” AMOS Conference, 2014.
[15] S. Han, et al., “Efficient generation of image chips for training deep learning algorithms,” SPIE Proceedings, 2017.
Date of Conference: September 19-22, 2023
Track: Satellite Characterization