Autonomous Detection and Classification of Lunar Minerals Using a Convolutional Neural Network Based Framework for the SUCR DALI Project

Raman spectroscopy is an important technology for planetary exploration with the potential to detect ice, minerals, and other important materials on various solid planetary bodies. This capability has already been demonstrated on the Martian surface through the perseverance rover [1]. The standoff ultracompact Ra-man (SUCR) instrument is one such technology, which is being developed for operation in a lunar setting through a funded Development and Advancement of Lunar Instrumentation (DALI) proposal, but has not yet become a flight ready instrument [2]. There is a need to implement functionality in such systems that can allow for real-time, autonomous detection of mineral, water, and other materials, depending on the objectives of the mission.
We developed a MATLAB-based CNN algorithm for automatically classifying minerals from Raman spectroscopy data. The current version of this algorithm is the Prototype 1 Raman Spectroscopy Autonomous Detection Classifier (ADC), which is part of the recently funded SUCR DALI proposal. This CNN prototype algorithm represents the first iteration in the ADC development roadmap and is designed to autonomously classify spectral data from samples containing primarily one type of mineral, but will eventually be extended to classify multiple minerals in mixture samples. Our results showed that this CNN-based algorithm, was able to successfully identify single minerals in spectra of mostly pure samples with a current validation accuracy of 92.86% as shown in Figure 1.