Quantitative Inversion of Lunar Surface Chemistry Based on Hyperspectral Feature Bands and Extremely Randomized Trees Algorithm

Abstract

In situ resource utilization (ISRU) is required for the operation of both medium and long-term exploration missions to provide metallic materials for the construction of lunar base infrastructure and H2O and O2 for life support. The study of the distribution of the lunar surface elements (Fe, Ti, Al, and Si) is the basis for the in situ utilization of mineral resources. With the arrival of the era of big data, the application of big data concepts and technical methods to lunar surface chemistry inversion has become an inevitable trend. This paper is guided by big data theory, and the Apollo 17 region and the area near the Copernicus crater are selected for analysis. The dimensionality of the first-order differential spectral features of lunar soil samples is reduced based on Pearson correlation analysis and the successive projections algorithm (SPA), and the extremely randomized trees (Extra-Trees) algorithm is applied to Chang’E-1 Interference Imaging Spectrometer (IIM) data to establish a prediction model for the lunar surface chemistry and generate FeO, TiO2, Al2O3, and SiO2 distribution maps. The results show that the optimum number of variables for FeO, TiO2, Al2O3, and SiO2 is 17, 5, 8, and 30, respectively. The accuracy of the Extra-Trees model using the best variables was improved over that of the original band model, with determination coefficients (R2) of 0.962, 0.944, 0.964, and 0.860 for FeO, TiO2, Al2O3, and SiO2, and root mean square errors (RMSEs) of 1.028, 0.672, 0.942, and 0.897, respectively. The modeling feature variables and model preference methods in this study can improve the inversion accuracy of chemical abundance to some extent, demonstrating the potential of IIM data in predicting chemical abundance and providing a good data basis for lunar geological evolution studies and ISRU.