Abstract
Acquiring information about the spatial distribution of ore grade in the subsurface is essential for exploring and discovering mineral resources. This information is derived commonly from the geochemical analysis carried out on drill core samples, which allows the quantification of the concentration of ore elements. However, these surveys are generally time-consuming and expensive, usually leading to information at a low spatial resolution due to large sampling intervals. The use of hyperspectral systems in the mining industry to characterize and quantify minerals in drill cores is increasing due to their efficiency and fast turnaround time. Here, we propose the use of convolutional neural networks on hyperspectral data to estimate Cu concentration in drill cores at the Olympic Dam iron oxide copper-gold deposit. The Cu concentration data obtained by drill core geochemical analysis and the mean spectra between the analyzed intervals obtained from hyperspectral data were used to train the machine learning model. The trained model was then used to estimate the Cu concentration of a test drill core, which was not used to train the model. In addition, the trained model was used to extrapolate the Cu concentration, at a centimetric spatial resolution, to the drill core intervals without geochemical analysis. Qualitative and quantitative evaluations of the results demonstrate the capabilities of the proposed method, which provided a root mean squared error of 0.48 for the estimation of Cu percentage along drill cores. The results indicate that the method could be beneficial for determining the spatial distribution of ore grade by supporting the selection of zones of interest where more detailed analyses are appropriate, reducing the number of samples needed to characterize and identify the ore zones, and assisting in the estimation of the volume with commercially viable ore, thereby potentially reducing the geochemical assays needed and decreasing the data acquisition time.
