Global Lunar Boulder Map From LRO NAC Optical Images Using Deep Learning: Implications for Regolith and Protolith

Abstract

Boulders on the lunar surface indicate relatively recent surface activity, related to mass wasting or bedrock excavation by impact cratering, and degrade over time, producing regolith. Previously, the distribution of boulders was indirectly assessed using the anisothermality effects observed by the Lunar Reconnaissance Orbiter (LRO) Diviner instrument. Here, we develop a pipeline based on a convolutional neural network to automatically identify and map individual boulders in LRO Narrow Angle Camera (NAC) images between 60°S and 60°N. Using $\sim$635,000 NAC images, we derive the first quasi-global inventory of lunar boulders consisting of about 94 million features with diameters larger than $4.5$ m. We determine relationships between crater diameter and sizes of ejecta boulders and find that the previously known higher boulder density in the mare regions relative to the highlands is due to a preferential location of boulders smaller than $\sim$10 m in the maria. The cumulative boulder size-frequency distributions (CSFDs) of simple crater ejecta are distinct between maria and highlands up to 130 m crater depth. This difference can likely be attributed to distinct subsurface rock contents, with a higher average mature regolith thickness in the highlands compared to the maria. Comparison of the derived boulder data set with the Diviner rock abundance map reveals broad, global agreement yet localized differences, attributable to different sensitivities of the two methods (optical images vs. thermal radiation) and variable geologic context. Diviner-NAC differences pinpoint to distinct lithologies, such as clast-rich zones and zones of fractured impact melt, typically extending for a few hundreds of meters laterally.