Geometallurgical estimation of mineral samples from hyperspectral images and topic modelling

Due to the development and consolidation of classical spectrographic techniques, many mining industries have extensive libraries of monopixel drill core spectra. In addition, laboratory characterization based on multispectral sampling by identifying and comparing reference spectral characteristics is used in many industrial processes. This work delves into the implementation of statistical generative modelling techniques for the characterization of mineral samples captured by hyperspectral imaging. The focus is on the generalization of previous developments based on hierarchical regression schemes on hyperspectral images for the modelling of geometallurgical properties that attempt to go beyond the direct identification of reference spectral features. The problem of characterizing the spatial-spectral variability of mineral samples is formalized as a topic modelling task, a central technique from the field of natural language processing. For this, we provide experimental evidence on how to organize hyperspectral pixels of a mineral sample for the definition of a corpus and the application of the popular topic modelling technique known as Latent Dirichlet Allocation (LDA). The use of LDA for the estimation of geometallurgical properties is demonstrated by presenting three ways of converting the spatial-spectral information to a corpus suitable for the LDA framework. A set of experiments was developed to validate the proposal. Two sample sets are presented for which laboratory characterization of geometallurgical properties and mineral occurrence measurements are available. These samples have also been captured using hyperspectral cameras in SWIR (Short-wave infrared) and VNIR (visible and near-infrared) wavelengths.

More information here