Abstract

Experiment results of neural network for predicting metal grades for gold, copper, and iron directly from the hyperspectral reflectance data of the drill core scans are presented. The technique is a combination of specialized pre-processing of the reflectance data, Convolutional Neural Networks (CNNs) and DenseNet, it was trained using data from Australian drill core scans and verified on Swedish drill core. A series of experiments were conducted to verify the robustness and adaptability of the model as well as the factors affecting the prediction accuracy of the Neural Network. Throughout the test, the model consistently demonstrated industrial-level prediction accuracy and robustness for different datasets. Moreover, the SWIR region of the hyperspectral reflectance proved to be a reliable dataset, while the LWIR region opened a wider range of applications in the exploration and mining operation. This paper provides details on a new way of using hyperspectral reflectance data directly to map metal grades and other targets. It also shows a way to get hyperspectral data easier usable for field geologists, mine operators and machine designers due to application of trained Neural Networks with limited processing efforts and a large robustness in the real world.

Keywords:Hyperspectral data analysis; Drill core; Base metals; Exploration; Artificial intelligence; Convolutional neural networks (CNNs)