Convolutional Neural Networks as a Tool for Raman Spectral Mineral Classification Under Low Signal, Dusty Mars Conditions

Abstract:

NASA’s Mars 2020 and ESA’s ExoMars missions will collect Raman measurements in field conditions that feature ubiquitous dust coverage obscuring underlying rocks and minerals. This presents a challenge for remote Raman measurements at distances where mechanical or ablative sample cleaning is not straightforward. Historically, only pristine targets tend to be sampled and/or low-quality spectra are discarded. We provide a means of identifying Raman spectra of common rock-forming silicate minerals under dusty, low signal to noise, Mars-like conditions using a convolutional neural network (CNN).

Fifteen minerals were identified including quartz, feldspars, amphiboles, micas, olivine, pyroxenes, calcite, gypsum, and mixtures of quartz, albite, forsterite, and augite. Pristine and dust-covered gabbro, of importance due to its compositional similarity to Mars’ basaltic crust, and granite rock spectra were used to test the CNN. Over 500,000 unique Raman spectra, with wide within-class variability, and 5000+ spectra per mineral class were acquired to train the CNN. Natural diversity in sample microtopography and crystallinity was used to generate varying laser focuses and spectral quality. No traditional spectral preprocessing was conducted, such as cosmic ray or baseline removal, and a Ricker wavelet transform was applied to standardize spectral datasets with varying resolutions and/or wavelength ranges.

Success scores of over 99% were achieved for pure mineral and mineral mixture identifications. The CNN effectively identified known low intensity Raman scatterers such as hornblende and biotite, and distinguished between mineral group end members such as albite, anorthite, and microcline. This is the first known implementation of true “big data” machine learning using varied, high-volume Raman spectral datasets, effectively tapping the capabilities of a CNN.

The pattern recognition abilities of CNNs have the potential to facilitate Raman spectral interpretation on Earth and Mars, increasing scientific yield, and even correcting human classification error. With the upcoming Mars missions, deploying surface rovers with onboard CNN classifiers would facilitate autonomous rover decision-making, enabling scientists to delegate increasingly complex tasks to the rover, saving time and resources.