Shane Ryall, Defense Research & Development Canada
Keywords: Image Processing, Deep Learning, SSA, NEOSSat
Abstract:
In this research, we investigate the application and deployment of utilizing deep learning object detection models for detecting satellites using imagery observed from space-based SSA platforms. Previous research in this field tends to use toy datasets produced through ground-based observations, which fails to account for specific difficulties faced by space-based assets or deployment in a full scale data processing pipeline. Some difficulties include differentiating cosmic rays from faint RSOs, and a frequent overlapping of background star streaks over the object. Our paper discusses the challenges associated with integrating modern deep-learning models into an operational SSA satellite data-processing pipeline, NEOSSat.
The paper will walk through the methodology employed for implementing a deep learning model for RSO detection, and how to integrate the model into the image processing loop for astrometric exploitation and orbit determination. We compare these methods against conventional RSO detection techniques which employ non-machine learning methods for RSO detection. This research is unique because it utilizes the largest space-based SSA dataset, derived from operations on the NEOSSat micro-satellite, we use a 30000 image dataset compiled from an ongoing Canadian GEO catalogue maintenance experiment. Additionally, the model must be trained in a resource-constrained environment as access to GPUs is limited. Given these constraints, we intend to show performance comparisons and highlight the potential pitfalls of deep learning for continuous real-world use.
Date of Conference: September 17-20, 2024
Track: Machine Learning for SDA Applications