RESUMO
Machine Learning (ML) has found several applications in spectroscopy, including recognizing minerals and estimating elemental composition. ML algorithms have been widely used on datasets from individual spectroscopy methods such as vibrational Raman scattering, reflective Visible-Near Infrared (VNIR), and Laser-Induced Breakdown Spectroscopy (LIBS). We firstly reviewed and tested several ML approaches to mineral classification from the existing literature, and identified a novel approach for using Deep Learning algorithms for mineral classification from Raman spectra, that outperform previous state-of-the-art methods. We then developed and evaluated a novel method for automatic mineral identification from combining measurements with two complementary spectroscopic methods using Convolutional Neural Networks (CNN) for Raman and VNIR, and cosine similarity for LIBS. Specifically, we evaluated fusing Raman + VNIR, Raman + LIBS or VNIR + LIBS spectra in order to classify minerals. ML methods applied to combined spectral methods presented here are shown to outperform the use of a single data source by a significant margin. Our approach was tested on both open access experimental Raman (RRUFF) and VNIR (USGS, RELAB, ECOSTRESS) libraries, as well as on synthetic LIBS (NIST) spectral libraries. Our cross-validation tests show that multi-method spectroscopy paired with ML paves the way towards rapid and accurate characterization of rocks and minerals. Future solutions combining Deep Learning Algorithms, together with data fusion from multi-method spectroscopy, could drastically increase the accuracy of automatic mineral recognition compared to existing approaches.
RESUMO
Future human missions to the surface of the Moon and Mars will involve scientific exploration requiring new support tools to enable rapid and high quality science decision-making. Here, we describe the PANGAEA (Planetary ANalogue Geological and Astrobiological Exercise for Astronauts) Mineralogical Database developed by ESA (European Space Agency): a catalog of petrographic and spectroscopic information on all currently known minerals identified on the Moon, Mars, and associated with meteorites. The catalog also includes minerals found in the analog field sites used for ESA's geology and astrobiology training course PANGAEA, to broaden the database coverage. The Mineralogical Database is composed of the Summary Catalog of Planetary Analog Minerals and of the Spectral Archive and is freely available in the public repository of ESA PANGAEA. The Summary Catalog provides essential descriptive information for each mineral, including name (based on the International Mineralogical Association recommendation), chemical formula, mineral group, surface abundance on planetary bodies, geological significance in the context of planetary exploration, number of collected VNIR and Raman spectra, likelihood of detection using different spectral methods, and bibliographic references evidencing their detection in extraterrestrial or terrestrial analog environments. The Spectral Archive provides a standard library for planetary in-situ human and robotic exploration covering Visual-Near-Infrared reflective (VNIR) and Raman spectroscopy (Raman). To populate this library, we collected VNIR and Raman spectra for mineral entries in the Summary Catalog from open-access archives and analyzed them to select the ones with the best spectral features. We also supplemented this collection with our own bespoke measurements. Additionally, we compiled the chemical compositions for all the minerals based on their empirical formula, to allow identification using the measured abundances provided by LIBS and XRF analytical instruments. When integrated into an operational support system like ESA's Electronic Fieldbook (EFB) system, the Mineralogical Database can be used as a real-time and autonomous decision support tool for sampling operations on the Moon, Mars and during astronaut geological field training. It provides both robust spectral libraries to support mineral identification from instrument outputs, and relevant contextualized information on detected minerals.
