Geochemical and Hyperspectral Data Fusion for Drill-core Mineral Mapping

Hyperspectral imaging is increasingly being used in the mining industry for the investigation of drill-core samples. It provides the means to analyze a large amount of cores considerably faster than traditional methods and in a non-invasive and non-destructive manner. Traditional approaches used to analyse drill-core hyperspectral data are mainly based on visual observations and need significant human interactions. Thus, they are time-consuming and subjective. In this paper, we explore the use of supervised machine learning techniques for mineral mapping in drill-core hyperspectral data. For this purpose, we suggest to use geochemical data for generating a training set. The main contribution of this work is to fuse geochemical and hyperspectral data within a machine learning framework. Moreover, for a more complete mineral mapping task, we integrate visible near-infrared (VNIR), short-wave infrared (SWIR) and long-wave infrared (LWIR) hyperspectral data. For the extraction of input features, the traditional Principal Component Analysis (PCA) is implemented. For classification, we propose to use Random Forest (RF) because of its significant performance in hyperspectral data classification when there are few training samples available. Experimental results show that the proposed method provides comprehensive mineral maps in which the distribution and patterns of different minerals are well characterised.