Stacey Jones, O Analytics, Inc.
Keywords: Super Resolution, low Earth Orbit, LEO, Satellite Detection, Object Characterization
Abstract:
Dependence on the Space Surveillance Network (SSN) or other monitoring networks that are comprised primarily of traditional sensing technologies complicates real-time object detection and classification. The quality of imagery emanating from these sources may be adversely affected by sensor age, physical location and/or orientation. Moreover, status quo artificial intelligence/machine learning (AI/ML) solutions are generally not well suited for high performance on blurred and otherwise degraded imagery, as they generally don’t bode well with obscure representations. We present a Super-Resolved Object Characterization in Low Earth Orbit (SROC LEO) approach that optimally exploits blurred imagery consistent with that made available from a significant number of SSN legacy sensor assets employing optical technology. By uniquely combining proven super-resolution and fast kinematic methods, early indications are that our SROC LEO concept will surpass performance of status-quo satellite classification to achieve improvement by a factor of four (4) in each of the two (2) pixel plane dimensions.
SROC LEO is a deep learning adaptation of a medical imaging solution to discover potential biomarkers (i.e., for Alzheimer’s diagnosis) using super-resolution, combined with an innovative target chipping method specifically tailored for satellite pixel region determination. The result is significantly improved near real-time space object characterization for low earth orbit (LEO) targets – specifically satellites. SROC LEO builds upon, extends and improves an existing GAN architecture proven to identify biomarkers, incorporates novel and efficient chipping preprocessing, and includes several computational and processing adaptations to yield superior space object characterization. Applying principles of Optical Coherence Tomography (OCT) retinal Generative Adversarial Network (GAN) architecture, and pixel-plane image processing techniques, in varying order based on image quality, lends to measurably surpassing performance of status-quo satellite classification.
SROC LEO deep satellite image detection, super-resolution, and classification consists of three core modules. The satellite detection module detects and extracts the region of interest (ROI) referred to as target (i.e., satellite) chips using novel algorithms that include edge detection (i.e., innovative target chipping (ITC)). These satellite chips undergo an adapted super-resolution series by an up-scale factor of 4/8 using a further adapted Multi-scale (MS) SRGAN super-resolution module followed by a deep VGG-based satellite classifier to initially identify the type of the satellite in the region of interest. The multi-stage super-resolution network is jointly trained with the satellite classifier. The final classification step involves correlating features extracted during ITC.
Discovery during development led an examination of performance gain by conditionally changing the order of execution of the core modules. Specifically, changing the order of the ITC and adapting super-resolution if/as needed to realize additional improvement. Also, in addition to superior performance, an important aspect of SROC LEO is execution in near real-time. As such, we pay special attention to execution times under different recognition scenarios. A comprehensive performance metric rubric is established to adequately evaluate SROC LEO performance in the field.
SROC LEO can significantly improve real-time space object characterization for low earth orbit (LEO) targets – specifically satellites; by uniquely combining proven super-resolution and fast kinematic methods. It is designed for use with current Space Surveillance Network (SSN) optical sensor technology and under conditions that involve noisy and grossly imperfect imagery. Successful SROC LEO deployment can be pivotal to improving space domain and situational awareness (SDA/SSA), rendezvous proximity operations (RPO), and in-space mission planning.
Date of Conference: September 17-20, 2024
Track: Satellite Characterization