Crowd science project for AI-based moon crater detection
Paul Nasdalack
Projektleitung: Prof. Stephan Jacob
Dauer: 12/2023 bis 11/2024
Förderer/Auftraggeber: European Space Agency (ESA)
ikum
Future lunar surface missions require extensive knowledge of hazards close to the landing site. The most prominent hazard type on the Moon are craters, existing in high numbers and varying in size and appearance. Automatically detecting craters is therefore a key technology for lunar landings. This applies to the a priori surface analysis and landing site selection as well as to the final descent and fully automated touchdown of the mission when the algorithm must assess landing safety independent from human interaction. While conventional algorithms face issues to detect these hazards in images with a sufficient success rate, software involving AI seems to be a very promising candidate to solve this challenge. However, this approach requires a large data set of fully labelled craters in lunar images to train the AI, which does not yet exist.
The objective of the proposed project is the involvement of the public to generate the AI training dataset. However, classic crowd science projects usually suffer from a rapid loss of interest by most participants due to the tedious nature of the tasks.This challenge is often tackled by including some ‘gamification’ elements like a scoring system to keep participants engaged. This project goes one step further: designing a game first and foremost, which is expected to draw a significantly larger user base. This would not only maximise the output of crater labels but also offers the potential to promote ESA’s innovation and to inspire the next generation of spaceflight engineers.
Following a feasibility study and a prototype development, the proposed project aims at developing and publishing a game that involves marking craters in surface images.
By playing the game, users will build up a training data set for an AI to automatically detect craters for future lunar surface missions.
If proven successful, the proposed approach has the potential to revolutionize the way AI training data is generated across various domains.