Evaluating the performance of hyperspectral short-wave infrared sensors for the pre-sorting of complex ores using machine learning methods

Abstract

Sensor-based sorting is increasingly used for the concentration of ores. To assess the sorting performance for a specific ore type, the raw materials industry currently conducts trial-and-error batch tests. In this study, a new methodology to assess the potential of hyperspectral visible to near-infrared (VNIR) and short-wave infrared (SWIR) sensors, combined with machine-learning routines to improve the sorting potential evaluation, is presented. The methodology is tested on two complex ores. The first is a tin ore in which cassiterite—the target mineral—is variable in grain size, heterogeneously distributed and has no diagnostic response in the VNIR-SWIR range of the electromagnetic spectrum. However, cassiterite is intimately associated with SWIR active minerals, such as chlorite and fluorite, which can be used as proxies for its presence. The second case study consists of a copper-gold porphyry, where copper occurs mainly in chalcopyrite, bornite, covellite and chalcocite, while gold is present as inclusions in the copper minerals and in pyrite. Machine-learning techniques such as Random Forest and Support Vector Machine applied to the hyperspectral data predict excellent sorting results in terms of grade and recovery. The approach can be adjusted to optimize sorting for a variety of ore types and thus could increase the attractivity of VNIR-SWIR sensor sorting in the minerals industry.