With investment in lunar exploration growing, the landscape is rich with opportunities to build faster, better, more reliable systems. This work focuses on the crucial yet under-studied problem of image-based novelty detection. Applications within this domain are far-reaching, already having shown potential to alleviate data transmission constraints, accelerate the pace of operations, and increase the overall science return of
exploration missions. As the field matures, novelty detection will play an essential role in the development of fully autonomous science platforms.

In the first part of this work, we advance the foundations of novelty detection for lunar exploration by developing a large, fully labelled, high-fidelity lunar analogue dataset. This dataset contains features and lighting conditions found across the Moon’s surface including fresh and degraded craters, rock fields, boulders, outcrops, and hills. It also contains various novelties including volcanic rocks, pyroclastic deposits, and exposed bedrock.

In the second part of this work, we establish new approaches to novelty detection for image data. Developing and experimenting with various models and datasets, we introduce adversarial autoencoders into the field of planetary novelty detection and build state-of-the-art variational and convolutional autoencoders, beating comparable state-of-the-art methods by >7%. In conjunction with the models themselves, we develop a framework to integrate region proposals with novelty detection to annotate novel loci
with bounding boxes. We also extend the scope of novelty scoring measures by operating directly on low-dimensional representations of image data.

In the final part of this work, we synthesize our findings into the context of lunar exploration and discuss key outcomes and future work.