Fast and Robust In-Situ Quantitative Mineral Analysis of Carbonate Rocks with AI-Driven Hyperspectral Unmixing Approach

Abstract Hyperspectral sensor measured the reflected electromagnetic wave within visible, near, middle, and long-wave infrared intervals. In geology, Short-wave Infrared (SWIR) and Long-wave Infrared (LWIR) wavelength ranges have been proven to be useful to detect mineralogy variation of a material. The purest material is usually unable to be captured in many cases, especially in distance measurements such as airborne or satellite. Each measurement, a single pixel in an image or point, usually captures the spectra from the mixture of several minerals. Spectral unmixing is a method to estimate the abundance of constituent minerals and their distribution in the mixed material of an object being measured. Using endmembers retrieved from the purest endmember measured or spectral libraries, the relative composition (abundance) of every material is quantified by analyzing the influence of the endmember's spectra on the noisy/impure spectra of the given mixed material. In this project, the abundance of Calcite, Dolomite, and Quartz of rock powder samples from carbonate outcrops was quantified using several techniques, including linear, PLSR, and machine learning assisted hyperspectral unmixing methods, and the results were compared with other conventional determination methods. The powder samples are measured using a handheld hyperspectral measurement device, a spectroradiometer, that has a 2 cm measured surface. XRD measurement was used to validate the mineral abundance quantification performed by every method applied to the measured spectra. The result of this study highlights the potential applications of a portable spectroradiometer as initial quantitative mineral analysis, leveraging the fast and non-destructive data acquisition supported by the AI unmixing algorithm for abundance estimation.