EFFECTS OF INCREASING THE DIMENSIONS OF CONVOLUTIONAL NEURAL NETWORKS (CNN) ON THEIR PERFORMANCE AND ACCURACY FOR ORE GRADE ESTIMATION FROM DRILL CORE HYPERSPECTRAL DATA
Master’s Dissertation
Record information
Authors
Karabo Elsie MAHLAPHA
Publisher
University of the Free State
Date
Collections
Description
The applications of deep learning for ore grade estimation from hyperspectral data are increasing. Currently, convolutional neural networks (CNN) are being used to extract deep features from data and learn relationships that exist between hyperspectral and ore grade data. However, these algorithms are still below optimal for ore grade estimation. This study endeavoured to investigate if the accuracy of ore grade estimation can be increased by changing the architecture of these algorithms and improve them for non-image hyperspectral data, for which CNN models are not widely applied. The dimensions of the CNN models were increased from 2D to 3D and the models were trained on spectrogram plots obtained from the short-time Fourier transform of reflectance spectra that were acquired from drill cores of the Olympic dam deposit in Australia. A ResNet-3D model achieved a root mean squared error (RMSE) of 0.238 ppm for Au grade estimation from data that had 75% of the Au grades below 0.625 ppm. This was lower than the 0.531 ppm RMSE achieved by a DenseNet-2D model on the same dataset and the 0.48 Cu% RMSE achieved by a VGG16 2D CNN on a dataset that had 75% of the data points below 1 Cu% reported in the literature. These results imply that the ResNet-3D model achieved higher accuracy of Au grade estimation for individual points within the data and that the under- or overestimation of the Au grades by this model would be lower than the other 2D CNN models when applied to testing data. Thus, the results imply that, in the application of CNN for ore grade estimation in research and in the mining industry, 3D CNN models should be considered. v