Unsupervised Deep Learning for Anomaly Detection on the Surface of the Moon

The Lunar Reconnaissance Orbiter (LRO) has been collecting high-resolution images (e.g. Narrow Angle Camera (NAC) images at ~0.5-2 m/pixel linearly) of the Moon since 2009, amassing a large dataset of images and offering researchers the opportunity to study the surface of the Moon and all that resides on it. This project aims to identify anomalous features across the Moon’s surface, locating not only scientifically useful natural formations such as rockfalls, fresh impact craters, irregular mare patches, or cave entrances, but also artificial objects such as landed spacecraft. To examine the millions of patches that make up the LRO surface mapping, we employ the use of a β-Variational Autoencoder (VAE) created by Lesnikowski et al. 2024 [1] and trained on ~850 NAC images (~52 million patches of 64 x 64 pixels) to perform this task. This overcomes the need for slow, by-hand labeling of the numerous images in the LRO database. These investigations will further gauge the model’s ability to locate anomalous surface features, both natural and artificial, at a statistically significant rate. Statistical metrics included the Kolmogorov-Smirnov (KS) two-sample test, Precision Recall (PR) curve analysis, Average Precision (AP) calculations, and the examination of both p-values and q-values (e.g. the model identified images of the Apollo 16 lander with a KS test value of 0.731, a p-value of 2.96e-9, and a q-value of 8.14e−9). A byproduct of these investigations is a database of ~270, 000 highly anomalous (> 3σ from the mean anomaly score of all the processed lunar images) features to be further examined by experts. The methodologies and products of this work may assist in locating lost or unknown technologies on the lunar surface and can aid in future crewed exploration of the Moon as major government organizations set their sights on returning to our nearest cosmic neighbor.