Assessing the reliability of an automated system for mineral identification using LWIR Hyperspectral Infrared imagery.

• This study investigates the reliability of a machine learning system to identify eleven different mineral grains. • A Long Wave Infrared (LWIR, 7.7–11.8 μm) imaging system with a macro lens used to improve spatial resolution. • Micro X-ray Fluorescence (µXRF) represented the mineral aggregates while Scanning Electron Microscope (SEM) showed the grains' surface. • The system accuracy was determined using two sets of ground truth (i.e. rigid-GT, and observed-GT). • The results of µXRF and the automated mineral identification system were compared using ArcGIS. Application of hyperspectral infrared imagery for mineral grain identification suffers from a lack of prediction on the irregular grain's surface along with the mineral aggregates. Here, we present an investigation to determine the reliability of automatic mineral identification in the longwave Infrared (LWIR, 7.7–11.8 μ m) with an LWIR-macro lens having a spatial resolution of 100 μ m. We attempt to identify eleven different mineral grains (biotite, epidote, goethite, diopside, smithsonite, tourmaline, kyanite, scheelite, pyrope, olivine, and quartz). A machine learning-based algorithm (implemented by software) compares all of the pixel-spectra to the ASTER spectral library of JPL/NASA using spectral angle mapper (SAM) and normalized cross-correlation (NCC) to create false-color maps. Then a hue-saturation-value (HSV) principle component analysis (PCA) based K-means clustering approach groups the mineral regions in different categories. The results were compared to two different ground truths (GT) (i.e. rigid-GT and observed-GT) for quantitative calculation and as an integrated step for validating our approach. Observed-GT increased the accuracy up to 1.5 times higher than rigid-GT, from 45.67 % to 69.39 %. The samples were also examined by micro X-ray fluorescence (μ XRF) and scanning electron microscope (SEM) in order to retrieve information on the mineral aggregates and the grain's surface. The results of μ XRF imagery (aggregate map) were compared to the results of automatic mineral identification techniques, using ArcGIS software, and the results represent a promising performance for automatic identification. [ABSTRACT FROM AUTHOR]