New candidate cave entrances on the Moon found using deep learning

Abstract

Pits and skylights are circular to elliptical, rimless, steep-sided depressions on planetary surfaces formed through gravitational collapse, which are of interest for astrobiological investigation and future space exploration. This is due to their ability to signify the presence of, or allow access to, underground cave systems such as lava tubes. The Lunar Pit Atlas contains 16 such features situated within mare regions that were partly discovered via the automated PitScan tool, which was limited by searchable latitudes and data coverage. In order to search for pits and skylights within these unmapped regions, we have trained a series of Mask R-CNN (Region-based Convolutional Neural Network) models on various combinations of Lunar and Martian remote-sensing imagery to detect Lunar pits and skylights. The best-performing model, named ESSA (Entrances to Sub-Surface Areas), was trained upon all available training data with a ResNet50 backbone. During testing on imagery of the famous Mare Tranquillitatis Pit and self-produced mosaics of proposed lava tube collapses, ESSA achieved average F${}_1$-scores of 82.4 and 93.7% for the bounding boxes and predicted masks, respectively. Despite only having surveyed $\approx$1.92% of the Lunar maria, ESSA has detected two previously uncatalogued skylights: the South Marius Hills and Bel’kovich A Pits (SMHP and BAP) - which are possible candidates for cave entrances on the Moon.