Unsupervised textural classification of rocks in large imagery datasets

Abstract

• Rock imaging characterisation techniques capture composition and texture. • A new feature extraction method for rock texture from imaging data is documented. • A large dataset of hyperspectral rock images was classified using this new method. • The results prove a robust and repeatable approach for textural analysis. • Textural outputs can be incorporated in geometallurgical modelling. The value of rock characterisation, whether it is for mineral exploration , extraction, or concentration, lies in the ability to describe its composition and texture. In the past century, extensive technological developments have provided new opportunities to assess compositional properties, both geochemical and mineralogical, and at different scales: X-ray diffraction, X-ray fluorescence, inductively coupled plasma mass spectrometry , reflectance spectroscopy , etc. More recently, the advent of imaging characterisation techniques and high-performance computing power has enabled the assessment of mineral texture in a robust and quantitative manner. This study proposes and validates an end-user focused workflow for the identification of textural families in a large drill-core hyperspectral imagery dataset, based on a novel textural feature extraction method named Mineral Co-Occurrence Probability Field (MCOPF). This workflow combines vintage image textural assessment methods with modern machine learning techniques for the automated unsupervised classification of textures within a drill core hyperspectral imagery dataset. The results demonstrate a meaningful and robust identification of rock textural families (clusters), enabling a wide range of applications in geology, mining, and metallurgy in the academic and industrial sectors.