New maps of major oxides and Mg # of the lunar surface from additional geochemical data of Chang'E-5 samples and KAGUYA multiband imager data

Abstract

In the past, global maps of major oxides and magnesium number (Mg #) on the lunar surface had been derived from spectral data of remote sensing images, combined with “ground truth” geochemical information from Apollo and Luna samples. These compositional maps provide insights into the chemical variations of different geologic units, revealing the regional and global geologic evolution. In this study, we produced new maps of five major oxides (i.e., Al2O3, CaO, FeO, MgO, and TiO2) and Mg # using imaging spectral data from the KAGUYA multiband imager (MI) and the one-dimensional convolutional neural network (1D-CNN) algorithm. We took advantage of recently acquired geochemical information from China's Chang'E-5 (CE-5) samples. We used the coefficients of determination (R2) and Root Mean Squared Error (RMSE) as model evaluation indicators. We compared the results with the previous machine learning algorithm models. Our study shows that the 1D-CNN algorithm model used in this study had a higher degree of fit and smaller dispersion between the “ground truth” value of geochemical information and the predicted value of spectral data. The 1D-CNN algorithm generally performs better in describing the complex nonlinear relationship between spectra and chemical components. In addition, we present regions of mare domes in Mairan Dome (43.76°N, 49.90°W) and irregular mare patches (IMPs) in Sosigenes (8.34°N, 19.07°E) to demonstrate the geologic implications of these new maps. With the highest spatial resolution (∼ 59 m/pixel), these new maps of five major oxides and Mg # will serve as an important guide in future studies of lunar geology.