AUTHOR=Stefanuk Braden , Skonieczny Krzysztof 

TITLE=Novelty detection in rover-based planetary surface images using autoencoders

JOURNAL=Frontiers in Robotics and AI

VOLUME=Volume 9 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/robotics-and-ai/articles/10.3389/frobt.2022.974397

DOI=10.3389/frobt.2022.974397

ISSN=2296-9144

ABSTRACT=In the domain of planetary science, novelty detection is gaining attention because of the operational opportunities it offers, including annotated data products and downlink prioritization. Autoencoders are known to be useful tools for novelty detection. This work improves upon state-of-the-art novelty detection performance in the context of martian exploration, by >7% on the metric of area under the receiver operating characteristic curve (ROC AUC), using a variational autoencoder (VAE). Autoencoders, and especially VAEs, perform well across all classes of novelties defined for martian exploration. VAEs are shown to have high recall in the martian context, making them particularly useful for on-ground processing of a large dataset of thumbnails. Convolutional autoencoders (CAEs), on the other hand, demonstrate high precision making them good candidates for onboard downlink prioritization applications. Adversarial autoencoders (AAEs) are on par with state-of-the-art in our implementation. Dimensionality reduction is a key feature of autoencoders for novelty detection, and VAEs and AAEs achieve comparable ROC AUC to CAEs despite observably poorer (highly blurred) image reconstructions; this is observed both in martian data and in lunar analogue data.