A Synthetic Image Data Generation Pipeline for Spacecraft Fly-by Scenarios

Accessible, high-quality datasets for machine learning activities remain limited to this date. Especially for space applications, this is relevant, but even more difficult is it for exploration missions, where the data to be obtained is not known beforehand. To reduce this lack of data, this work proposes a pipeline for automatic realistic data generation for fly-by scenarios such as planned for ESA’s Comet Interceptor mission. The pipeline consists of two components. First, creating digital models of the nucleus and dust cloud using Blender and second, adding camera artefacts with Python. The camera artefacts considered here are radiation effects, image fractures, pixel errors, thermal noise, and dark current. The goal of the pipeline is to generate realistic data with camera artefacts to supplement the limited available real image footage. Additionally, the synthetic generation of data with random combinations of factors will allow to cover many possible scenarios, for example, irregular body shapes like was observed during ESA’s Rosetta mission. The generated data will be used to train efficient neural networks to be deployed on-board spacecrafts’ future processing systems which enable increased data reduction capabilities in space. The pipeline is also designed to create data to verify the system’s behaviour in certain scenarios.