Insights into Lunar Mineralogy: An Unsupervised Approach for Clustering of the Moon Mineral Mapper (M3) spectral data

Abstract

This paper presents a novel method for mapping spectral features of the Moon using machine learning-based clustering of hyperspectral data from the Moon Mineral Mapper (M3) imaging spectrometer. The method uses a convolutional variational autoencoder to reduce the dimensionality of the spectral data and extract features of the spectra. Then, a k-means algorithm is applied to cluster the latent variables into five distinct groups, corresponding to dominant spectral features, which are related to the mineral composition of the Moon's surface. The resulting global spectral cluster map shows the distribution of the five clusters on the Moon, which consist of a mixture of, among others, plagioclase, pyroxene, olivine, and Fe-bearing minerals across the Moon's surface. The clusters are compared to the mineral maps from the Kaguya mission, which showed that the locations of the clusters overlap with the locations of high wt% of minerals such as plagioclase, clinopyroxene, and olivine. The paper demonstrates the usefulness of unbiased unsupervised learning for lunar mineral exploration and provides a comprehensive analysis of lunar mineralogy.

Figures (10)

• Figure 1: Outline of the three major terranes as defined by Jolliff2000, with the figure taken from MARTINOT2020113747. The PKT is in blue, the SPAT is in red, and the FHT-a is in yellow. The surface that is not comprised in the PKT, SPAT, or FHT-a is the FHT-o.
• Figure 2: The Moon Mineralogy Mapper (M3) reflection data of the Moon. The orange squares highlight the areas of the dataset used to train the convolutional variational autoencoder. The upper map is displayed in an equidistant cylindrical projection with a maximum latitude of 75 degrees. The bottom left and right show the Moon's south and north poles, respectively, in a stereographic projection with a minimum of 60 degrees.
• Figure 3: Overview of the process of training the neural network and clustering of data. The spectra are passed through a symmetric convolutional bottleneck autoencoder and reconstructed. The loss between the original spectra and the reconstructed spectra is minimized. The encoded latent variables contain the relevant features needed to reconstruct the spectra. Therefore, the latent variables are used to represent the spectra in low-dimensional space. After the latent variables are obtained for each pixel, they are clustered by k-means.
• Figure 4: Structure of the variational autoencoder. The input has the form of the number of wavelengths. The convolutional layers decrease the features of the data while increasing the number of kernels in the convolutional layers. E.g. in the encoder, the data is more compressed going from left to right. After the convolutional layers, the features are flattened before going into the bottleneck consisting of a linear layer. The decoder is the inverse of the encoder, consisting of linear layers, with the sigmoid activation function at last.
• Figure 5: The resulting Moon spectral cluster map from the deep learning clustering approach. The top is in equirectangular projection from -70 to 70 degrees in latitude. The bottom left is the south pole in stereographic projection from -60 degrees latitude, and the bottom right is the north pole in stereographic projection from 60 degrees latitude. The black regions have zero reflectance, and the white regions have no data.