Performance results of predictive models for estimating chemistry grades for gold, copper and iron in drill cores, based on the mineralogy data derived from the hyperspectral observations and using automated tools are presented. The models were built using data from more than 700 km of drill core. Ten commonly used machine learning and neural network algorithms, including convolutional neural networks (CNNs), were assessed for classifying ore grades, with accuracy and errors reported through confusion matrixes. The CNN algorithm was the outstanding performer, with an averaged classifier accuracy of 80%, outperforming the other machine-learning methods and the DenseNet deep learning method. Also discussed is the outcome of using fewer ore-grade classes that led to better predictive accuracy. This work provides insight into the potential for predicting geochemistry from hyperspectral data to support exploration geologists in target detection.

Hyperspectral infrared remote sensing has become an extremely useful and increasingly available tool for mineral exploration and geoscience in general over the last 40 years.¹ A major goal of hyperspectral remote sensing has been the quantification of mineral abundance, not only for mineral exploration and geologic mapping on this planet but also for mineral mapping on other planets. Phyllosilicates, and specifically clay minerals, are a key mineral group of interest to geologists as this group of finely crystalline phyllosilicates is known to trace hydrothermal and weathering processes relevant for understanding mineralization processes and geometries.

This paper summarises an evaluation of the application of artificial intelligence to hyperspectral drill-core scans for more effective mineral exploration. The dataset used was based on publicly available core scans and related geochemical analysis from Australia. Prior to unification, a detailed quality assessment of the geochemical data was undertaken. Special focus was paid to gold, silver, copper, iron, uranium, nickel, lead, tin, antimony, arsenic and bismuth contents. The dataset was labelled with defined ore grades related to economic cutoff values. The impact on predictions of different setups is related to the amounts of data used for learning, data design and implementation of the geological domains. Based on 1-metre bins, the results from more than 700 km of drill cores were used and analysed with the potential for geological exploration in different scenarios discussed. The results indicate the enormous potential of the use of hyperspectral scans in combination with artificial intelligence for the development of exploration scenarios and to provide support for exploration geologists and target detection. The application of predictors on scanned drill cores from Australia also indicates mineralised zones that have not been analysed chemically for all metals above economic cutoffs. This result shows the enormous potential of the approach for strategic exploration but also mining operations. Prediction of geochemical concentrations for gold, copper and iron based on a neural network in drill cores is possible. Using mineral abundances from hyperspectral core scans as learning records, and existing elemental geochemical analyses as labels, the predictions are given with an accuracy of better than 80–90%.

Reflectance spectroscopy is a nondestructive, rapid, and easy-to-use technique which can be used to assess the composition of rocks qualitatively or quantitatively. Although it is a powerful tool, it has its limitations especially when it comes to measurements of rocks with a phaneritic texture. The external variability is reflected only in spectroscopy and not in the chemical-mineralogical measurements that are performed on crushed rock in certified laboratories. Hence, the spectral variability of the surface of an uncrushed rock will, in most cases, be higher than the internal chemical-mineralogical variability, which may impair statistical models built on field measurements. For this reason, studying ore-bearing rocks and evaluating their spectral variability in different scales is an important procedure to better understand the factors that may influence the qualitative and quantitative analysis of the rocks. The objectives are to quantify the spectral variability of three types of altered granodiorite using well-established statistical methods with an upscaling approach. With this approach, the samples were measured in the laboratory under supervised ambient conditions and in the field under semisupervised conditions. This study further aims to conclude which statistical method provides the best practical and accurate classification for use in future studies. Our results showed that all statistical methods enable the separation of the rock types, although two types of rocks have exhibited almost identical spectra. Furthermore, the statistical methods that supplied the most significant results for classification purposes were principal component analysis combined with k-nearest neighbor with a classification accuracy for laboratory and field measurements of 68.1 and 100, respectively.