Automated Impact Melt Fracture Mapping on the Moon With Weakly Supervised Deep Learning

Abstract Cooling fractures found within impact melt deposits have been manually mapped within several craters on the Moon and Mercury, as their distribution can indicate which heat‐loss processes were most significant in the periods after impact. However, due to the discovery of melt deposits in Lunar impact craters with sub‐km diameters, it is unlikely that the complete mapping of these impact melt fractures (IMFs) on the Moon will be achievable without automation. As such, we have trained a DeepLabV3 semantic segmentation deep convolutional neural network, called IMFMapper, to detect IMFs within Lunar Reconnaissance Orbiter Narrow Angle Camera (LROC NAC) satellite imagery. As a means of maximizing the size of the training data set, “weak” pixel‐level labels were generated by buffering line annotations. In testing upon the IMFs found within Ohm crater, IMFMapper achieved an average ‐score of 69.3%. IMFMapper has also been deployed to map IMFs within the previously surveyed Crookes crater, where we have found new candidate melt deposits within the crater's western and southern walls. In addition, IMFMapper has produced the first map of IMFs within Schomberger A crater, in which IMFs may act as permanently shadowed regions due to the crater's proximity to the Lunar South Pole. The successful mapping of IMFs in Schomberger A also signifies IMFMapper's robustness to extreme solar incidence angles. We also demonstrate that IMFMapper could be implemented for automated mapping of IMFs on Mercury upon the commencement of BepiColombo's science operations.