Douglas Woodward, The Aerospace Corporation; Celeste Manughian-Peter, The Aerospace Corporation; Tim Smith, The Aerospace Corporation; Elizabeth Davison, The Aerospace Corporation
Keywords: Machine Learning, Space Domain Awareness
Abstract:
As space becomes increasingly congested across all orbits, Space Domain Awareness (SDA), requires ever more accurate detection, cataloging, tracking, predicting, and contextualizing of all space objects and activity.
A fundamental challenge in SDA is source extraction: identifying which pixels in an image contain Resident Space Objects (RSOs). For RSOs in Geosynchronous Earth Orbit (GEO), ground based, rate-tracked, electro-optical imagery is the primary data used for source extraction as radar techniques become less effective at higher orbits. However, there are limitations in applying this data type as the basis for source extraction of RSOs in GEO, primarily the distance between target and observation, atmospheric noise, background stars, low signal to noise ratios and high apparent visual magnitude. These limitations render manual source extraction difficult and, often, prohibitively time consuming, leading to a need for efficient, accurate automated alternatives.
Current automated source extraction approaches include intensity thresholding, peak detection, gaussian processes, connected components, RANSAC, frame stacking, manually tuned convolutional filters, wavelet filtering, and Bayesian methods. Most recently, convolutional neural networks (CNNs) have been applied as a method for object detection. Though their performance has been shown to be state of the art, object detection CNNs require large numbers of parameters and do not provide pixelwise labels.
Here, we contribute a novel CNN-based approach for the detection of GEO RSOs from ground-based, rate-tracked, electro-optical telescope imagery through image segmentation. We compare this approach with alternative deep learning approaches and demonstrate state of the art performance at lower pixel error thresholds with approximately 1% of parameters. This technique allows for pixel-wise labeling of an image, expanding possibilities for follow-on work in the field such as automated RSO breakup and collision detection, debris tracking, and closely spaced object detection.
Approved for public release. OTR 2021-00420.
Date of Conference: September 14-17, 2021
Track: Machine Learning for SSA Applications