Applying machine learning to a nonlinear spectral mixing model for mapping lunar soils composition using CHANDRAYAAN-1 M3 data

Abstract

We present a newly developed method which combines the nonlinear spectral mixing model of Shkuratov et al. (1999) with a machine learning algorithm to map the lunar regolith composition using spectral data. The new method performs orders of magnitude faster than the traditionally used numerical optimization approaches, allowing the mapping of regolith properties (including mineralogical composition, average grain size and optical maturity) over large areas of the lunar surface. A new set of basic mineral spectra of the lunar soil for using with spectral mixing models is proposed. Used together with the nonlinear mixing model (Shkuratov et al., 1999), the set is able to describes Chandrayaan-1 M3 instrument spectra collected from test areas which includes the Shapley crater with its surroundings containing mare and highland terrains well. The new set includes a virtual “gray component” with a “flat” (constant) spectrum, accounting for the factors that change general surface albedo, such as spectrally neutral components (e.g., agglutinate glasses), errors in the photometric reduction, uncertainties in estimations of lunar regolith porosity q and the mean grain size S of the basic minerals. The proposed new method takes into account the influence of space weathering and nonlinear correlation between the compositional and spectral parameters of the lunar soils delivering values for the optical properties and mineralogical abundance determination of the lunar regolith which are compatible with the results found from lunar samples measurements in the laboratory. The proposed approach can be used for analyzing spectral observations not only of the lunar surface but also for other surfaces with are covered by regolith.