Deep learning lunar penetrating radar inversion: An example from Chang’E-3

Abstract

The Moon’s deeper subsurface layers beyond the regolith are not well-studied. Using data from Change’E-3 Yutu rover’s lunar penetrating radar (LPR), we invert for its subsurface dielectric permittivity (ε_r) model. We use convolutional neural network based deep learning architecture to train numerous ε_r profiles and their corresponding synthetic radargrams. The dielectric permittivity training dataset is designed to encapsulate all possible ε_r realizations that the lunar subsurface materials may have. We test our trained model on synthetic data, and on the Change’E-3 LPR data. We validate our predicted ε_r by comparing its forward data and the field data. Our interpretation suggests multiple layers in the upper 200 meters in the order of regolith, ejectas, Eratosthenian basaltic lava flows, paleoregolith, and lava flows from Imbrium period.