Leon Muratov, Spectral Sciences Inc; Timothy Perkins, Spectral Sciences Inc; Marsha Fox, Spectral Sciences Inc; Xuemin Jin, Northeastern Universith ; Paul LeVan, Air Force Research Laboratory/RDST
Keywords: Satellite, 3D Reconstruction, Neural Network, Image, Model, QUID, Synthetic Image, Deep Learning, Artificial Intelligence.
Abstract:
With the growing number of inactive and active man-made space objects orbiting Earth, space situational awareness (SSA) is becoming an ever more critical element of for both national security and the economic use of space. Accurate tracking and identification of these resident space objects (RSO) through space-based or ground-based imagery can provide a means to characterize potential threats to our orbital assets, and to possibly infer the intent of foreign satellites. This work discusses the development of a new satellite identification tool that employs physical model predictions and deep learning neural networks (NN) to increase the quantity of information that can be extracted from these sources of space surveillance imagery.
When observing distant objects in geo-stationary or midcourse orbits, or small satellites in low orbits, unresolved imagery with less than a dozen resolution elements may be the only available optical measurements. Our satellite identification method, the Cognitive Image Recovery Code (CIRC), is designed to evaluate low resolution imagery, including views from multiple locations and temporal light curves generated from unresolved imagery, and to predict the most likely RSO configuration from a list of possible models. For a number of viewing and illumination angles, the tool applies a dense neural network to identify characteristic features of each model and evaluates the most likely match to an unresolved observation. The training imagery for the machine learning algorithm was generated using QUID (QUick Image Display), our fast, first-principles signature simulation code, and a catalog of physically attributed 3D models of various satellites and RSO types. The fidelity of these simulations ensures that the training imagery is both realistic and radiometrically accurate, and the computation speed generates images on-the-fly, allowing an iterative refinement of the model prediction.
A demonstration of the CIRC identification method has been conducted using a limited set of viewing configurations and RSO models. An evaluation of the performance of the tool against various levels of sensor noise and image resolutions is presented, and an assessment of the estimation uncertainties and their mitigations are discussed. Future development, including generalizations for unknown viewing geometry and automated model augmentation, are also discussed.
Date of Conference: September 17-20, 2019
Track: Machine Learning for SSA Applications