A Machine Learning Technique for Drill Core Hyperspectral Data analysis

Hyperspectral data are increasingly being used to map minerals in drill core samples allowing a non-invasive and non-destructive characterization of the mineral assemblages, and therefore, the mineralogical composition of a system, its variability, and structural features. The analysis of drill core hyperspectral data is traditionally carried out by a visual interpretation of the spectra and a comparison with reference libraries using spectral similarity measures. Although this approach produces good results it is time-consuming and subjective. In this work, we introduce, for the first time, an innovative automatic mineral mapping technique for drill core hyperspectral data by using a machine learning approach. More specifically, we propose to exploit detailed information coming from the Scanning Electron Microscopy (SEM)-based Mineral Liberation Analysis (MLA) to train a supervised classifier. For the extraction of input features, a traditional technique is explored, i.e., Principal Component Analysis (PCA). For the classification step, we suggest to use Random Forest (RF) because of its significant performance when there are few training samples available. Experimental results conducted on a VNIR-SWIR drill core hyperspectral dataset, show accurate classification results.