Global mapping of lunar refractory elements: multivariate regression vs. machine learning

Abstract

Context.The quantitative estimation of elemental concentrations at the spatial resolution of hyperspectral near-infrared (NIR) images of the lunar surface is an important tool for understanding the processes relevant for the origin and evolution of the Moon.Aims.We aim to map the abundances of the elements Fe, Ca, and Mg at a typical accuracy of about 1 wt.% at the spatial resolution of the Moon Mineralogy Mapper (M3) instrument on-board Chandrayaan-1 lunar mission.Methods.The NIR reflectance of the lunar regolith is an integrated response to the presence of refractory elements and soil alteration processes. Our approach was to define a combination of spectral parameters that are robust with respect to the effects of soil maturity. We calibrated the spectral parameters with respect to elemental abundances measured by the Lunar Prospector Gamma Ray Spectrometer (LP GRS) and the Kaguya GRS (KGRS). For this purpose, we compared a classical multivariate linear regression (MLR) approach and the machine learning based support vector regression (SVR) technique applied to M3global observations.Results.The M3-based global elemental maps are consistent in distribution and range with the LP GRS and KGRS elemental maps and do not show artifacts in immature areas such as small fresh craters. The results derived using MLR and SVR are compared to sample-based ground truth data of the Apollo and Luna sample-return sites, where the root-mean-square deviations obtained by the two regression models are similar.Conclusions.The main advantage of the proposed new algorithm is its ability to minimize artifacts due to space-weathering effects. The elemental maps of Mg and Ca provide additional information and reveal structures not always visible in the Fe map. The global elemental abundance maps derived for the fully calibrated M3observations might thus serve as important tools to investigate the lunar geology and evolution.