Novelty detection in rover-based planetary surface images using autoencoders

Area of Science:

• Planetary Science
• Artificial Intelligence
• Machine Learning

Background:

• Novelty detection is crucial for planetary exploration, enabling annotated data products and downlink prioritization.
• Autoencoders, particularly variational autoencoders (VAEs), are effective for identifying novelties in complex datasets.

Purpose of the Study:

• To improve state-of-the-art novelty detection in Martian exploration using VAEs.
• To evaluate the performance of different autoencoder architectures (VAEs, CAEs, AAEs) for novelty detection.
• To investigate the impact of dimensionality reduction on detection quality.

Main Methods:

• Implementation of variational autoencoders (VAEs) for novelty detection.
• Comparison of VAEs, convolutional autoencoders (CAEs), and adversarial autoencoders (AAEs).
• Analysis of detection performance using the area under the receiver operating characteristic curve (ROC AUC).

Main Results:

• VAEs demonstrate improved novelty detection performance in Martian exploration, achieving high recall.
• CAEs exhibit high precision, suitable for onboard downlink prioritization.
• AAEs perform comparably to state-of-the-art methods.
• Dimensionality reduction via VAEs and AAEs yields competitive ROC AUCs, even with less precise image reconstructions.

Conclusions:

• VAEs are highly effective for on-ground novelty detection in planetary science due to their high recall.
• CAEs are well-suited for onboard data prioritization owing to their high precision.
• Autoencoder-based dimensionality reduction is a viable strategy for robust novelty detection in planetary data.