Automatic Classification of Planetary Surfaces Using Autoencoders With Applications to Dunes in the North Polar Erg of Mars

The study of planetary surface processes has traditionally relied on the manual interpretation of spacecraft images. While manual image analysis methods are robust and well‐established, they become impractical when the volume of available data is large and may introduce observer bias. In recent years, supervised machine learning models have become common in augmenting human analysis of planetary surface processes. However, those models still require training on large, manually labeled data sets. In this work, we demonstrate the potential of unsupervised machine learning for planetary geomorphology by focusing on a morphological analysis of dunes near the northern polar erg of Mars. We employ two types of unsupervised neural networks, a convolutional autoencoder and a masked autoencoder, to obtain a compact representation of spacecraft images of Mars' northern‐polar region. The latent space features extracted by the models reveal the variety of terrain morphologies prevalent in the northern polar region of Mars: continuous barchanoid ridges, isolated barchan dunes, dark sediment, and regolith. Our findings, which match previous manual mapping efforts, show that autoencoders are useful for extracting geomorphological insights without the need for labeled data, providing a foundation for an automated and scalable way to study planetary surfaces.