Duration: 36 months
The most prominent hazards on the lunar surface are craters and boulders. The hazards need to be reliably detected, e.g., to be able to perform avoidance maneuvers, or for using them as orientation points. Thus, reliable detection is crucial to mission success. While conventional algorithms face difficulties to get sufficient success rates, modern cutting-edge AI approaches are very promising to solve this challenge. In this project, we develop such solutions and corresponding software (available to the ESA community) that enables automatic hazard detection in lunar surface data, thus contributing to enhance Europe’s lunar surface analysis capabilities and pave the way towards future lunar exploration. We start to investigate three main directions, and, after evaluation, focus on further developing the most promising one: (i) Unsupervised learning approach. It learns a normal model based on an unlabeled dataset. Deviation from the normal model allows to detect and localize the hazard. For its implementation, recent Student-Teacher strategies where a student network is trained to reproduce the features of a teacher network for normal samples are promising. (ii) A problem-adapted semantic segmentation approach. In contrast to classical object detection, we expect to be able to detect objects of smaller size. Further, we plan to incorporate causality into the machine-learning framework. This is very promising in order to integrate prior information and hence improving the success rate and ensuring the explicability of the decision-making. (iii) Zero-Shot Detection-based approaches. This novel breakthrough technology combines image and language features based on cross-modality deep learning architectures. With this type of transformers-based networks, detecting new classes (such as craters or any other new object) without retraining the model becomes possible. This is very promising in view of evolving cognitive capabilities in a long-term development strategy.
