Machine learning approaches for classifying lunar soils

We examine the ability of machine learning (ML) techniques to determine the physical and mineralogical properties of lunar soil using reflectance spectra. We use the Lunar Soil Characterization Consortium (LSCC) dataset to train and assess the predictive power of classification models based on their type (Mare soil and Highland soil), particle size, maturity, and the dominant type of pyroxene (High-Ca and Low-Ca). Nine ML algorithms including linear methods, non-linear methods, and rule-based methods (three from each) were selected, representing a range of characteristics such as simplicity, flexibility, computational complexity, and interpretability along with their ability to handle different types of data. Fifteen spectral parameters were initially introduced to the models as input features and a maximum of four features was selected as the best feature combinations to classify lunar soils based on their types, particle size, maturity, and the type of pyroxene. The Support Vector Machine with radial basis function (svmRadial) and the penalized logistic regression model (gitnnet) performed well for all target variables with high accuracies. Band depths and Integrated band depths at 1 mu m, 1.25 mu m and 2 mu m, band position of the 1 mu m band, along with four band ratios (band tilt, band strength, band curvature, and olivine/pyroxene) are important features for classifying soil type, grain size, maturity, and type of pyroxene from reflectance spectra. This study shows that proper preprocessing and feature engineering techniques are crucial for high performance of the predictive models.