Automated Discovery of Anomalous Features in Ultralarge Planetary Remote-Sensing Datasets Using Variational Autoencoders

The NASA Lunar Reconnaissance Orbiter (LRO) has returned petabytes of lunar high spatial resolution surface imagery over the past decade, impractical for humans to fully review manually. Here, we develop an automated method using a deep generative visual model that rapidly retrieves scientifically interesting examples of LRO surface imagery representing the first planetary image anomaly detector. We give quantitative experimental evidence that our method preferentially retrieves anomalous samples such as notable geological features and known human landing and spacecraft crash sites. Our method addresses a major capability gap in planetary science and presents a novel way to unlock insights hidden in ever-increasing remote-sensing data archives, with numerous applications to other science domains.