Deep learning-based extraction of dielectric constants of shallow lunar soil

The Yutu rover carried by the Chang'E-3 lander is equipped with the first lunar penetrating radar system(LPR), which is designed to detect the dielectric properties and the structure of the shallow Moon. Aiming at the limitations of traditional radar data analysis and manual feature extraction processes, this paper proposes a dielectric constant inversion network based on GPRNet. The network adds a differential layer and a self-attention mechanism to the base model, can predict the lunar soil subsurface structure directly using LPR data, and can effectively deal with the subsurface interface at different depths. The model performs well on one-dimensional simulation data, with the R2 metric improved to 0.9540, the RMSE reduced by 5%, and the depth error reduced by 27%. It demonstrates excellent generalization ability on data of different depths, and finally, the model is applied to the Chang'E-3 LPR dataset. The experimental results show that the method can effectively invert the corresponding dielectric constants based on LPR data to obtain the dielectric properties of lunar soil and buried materials.