Analysis of Lunar Rockfalls

Past exploration missions have revealed that the lunar topography is eroded through mass wasting processes such as rockfalls, similar to Earth. Despite the abundance of remote sensing and in–situ data many questions about the occurrence and characteristics of lunar rockfalls remain unanswered: Why do lunar rockfalls occur? What endo- and exogenic processes drive lunar rockfalls? What can these processes tell us about the geologic activity and evolution of the Moon? What can rockfall tracks teach us about the shallow lunar regolith and its interaction with the space environment? To help address these questions, we develop deep learning-driven algorithms that allow us to automate the detection and mapping of rockfalls in Lunar Reconnaissance Orbiter’s Petabyte-scale image archive - a uniquely rich resource that has barely been exploited by previous work. Our convolutional neural networks achieve human-level detection performance while reducing the required processing time by several orders of magnitude. By applying transfer learning principles we further boost the performance of our detectors and make them robust enough to map rockfalls on other bodies across the Solar System as well, such as Mars, Ceres, and comet 67P. Using such multi-domain trained algorithms we conduct the first-ever, global scan of the lunar surface and derive the global distribution of rockfalls on the Moon, identifying more than 130,000 individual rockfalls. The global distribution of rockfalls is highly heterogeneous, where the vast majority as well as the highest spatial concentrations of rockfalls occur in young impact craters and on equatorfacing slopes with large thermal gradients. This indicates that rockfall occurrence is mainly controlled by erratic impacts and continuous solar-driven thermal fatigue. We discover that rockfalls occur even in the oldest, pre-Nectarian lunar terranes, indicating that the identified processes - particularly small-scale meteoritic impacts - drive erosion over billions of years. We do not observe significant clusters of rockfalls in the proximity of visible volcanic and tectonic features which suggests that seismic activity has not been a major, global driver of lunar rockfall in the Moon’s recent geologic past. Using satellite images of rockfall boulders and their tracks, we further develop a method to estimate the regolith surface strength and geomechanical properties of yet unexplored regions, such as the polar permanently shadowed regions and large pyroclastic deposits, directly supporting humanity’s current efforts to continue the exploration of the Moon and beyond. This dissertation demonstrates the enormous potential of deep learning-driven applications for planetary science and reveals new insights about the weathering, erosion and evolution of airless planetary bodies in our solar system