Fast automated mineral differentiation using optical hyperspectral imaging

Abstract

The processing of mineral-bearing rocks in the mining industry is a highly energy and resource-intensive process. This has motivated the development of more effective ways to automatically sort valuable ore from barren material before the most expensive crushing and processing stages.

Many different sensor-based methods to perform automated ore sorting have been investigated in the literature and in the industry. The present work explores potential improvements to sensor-based sorting methods based on optical reflectance hyperspectral imaging.

These improvements are drawn from the field of remote sensing and revolve around the decomposition of reflectance spectral images into relatively small numbers of component spectra ("endmembers") and their corresponding abundance values. Classification models using linear discriminant analysis (LDA) based on these abundances are fitted and compared against classifiers operating directly on the original spectral data or on RGB colour values. All classifiers were fitted and evaluated on lithium-bearing pegmatite (lepidolite) samples to be differentiated from waste rock (weathering granite) from Gon?alo, Portugal.

The resulting abundance-based classifiers are found to be competitive and generalise well from training data to test data without risk of overfitting, unlike direct spectral classification models. They also offer the option of manually overriding the classification significance of individual spectral components, which is demonstrated on the example of contamination consisting of labels written on the surface of the samples.