A synthetic comet and asteroid image dataset for neural network training and system verification

Paper number

IAC-24,A3,IP,15,x90222

Author

Mr. Ric Dengel, University of Tartu, Estonia

Coauthor

Dr. Mihkel Pajusalu, University of Tartu, Estonia

Coauthor

Prof. Rene Laufer, Luleå University of Technology, Sweden

Year

2024

Abstract

In the realm of machine learning applications, access to high-quality datasets remains a challenge, particularly in the domain of space exploration where data acquisition is often unpredictable. This paper introduces a dataset created by FlyByGen, which is a novel pipeline designed to address this gap by generating synthetic comet and asteroid image datasets tailored for machine learning tasks and system verification of space missions.
FlyByGen is specifically developed to produce realistic datasets for fly-by scenarios, such as those encountered in missions like ESA's Comet Interceptor. Leveraging Blender and Python, the pipeline enables the generation of a diverse array of synthetic comet and asteroid images, encompassing various camera artifacts crucial for space imagery, including radiation effects, image fractures, sensor faults, thermal noise, and dark current.
The generated datasets serve as invaluable resources for training neural networks aimed at enhancing onboard spacecraft data reduction capabilities. Beyond just the raw data it allows to save the Blender files, but most importantly also ground truth masks, which are needed for the neural network training. Moreover, FlyByGen lays the groundwork for an end-to-end workflow, spanning from data generation to neural network training and eventual deployment on future FPGA platforms onboard spacecraft. Most importantly this dataset allows to create various proof-of-concept neural networks which allow to assess the processing complexity on space relevant hardware.
While the comet and asteroid dataset represents the initial focus of FlyByGen, the pipeline is designed to accommodate the generation of diverse datasets to verify system behavior in other space scenarios, including rendezvous and docking maneuvers and space situational awareness tasks. Through its versatility and adaptability, FlyByGen allows to create many various datasets and emerges as a pivotal tool for advancing machine learning applications and system verification, but also promotional videos in the realm of space exploration and specifically fly-by scenarios.

Abstract document

IAC-24,A3,IP,15,x90222.brief.pdf

Manuscript document

IAC-24,A3,IP,15,x90222.pdf (🔒 authorized access only).

To get the manuscript, please contact IAF Secretariat.