Multi-dimensional convolutional neural network application to hyperspectral reflectance data: A complementary tool for ore grade estimation

Abstract

Ore grade estimation research is evolving from applications of geostatistical methods to applications of AI models. One of the AI algorithms that is currently being researched for ore grade estimation is convolutional neural network (CNN). It has been found to achieve ore grade estimation, even for structurally-controlled deposits; however, its application still requires improvement to perform more accurately for this important task. In this study, the application of CNN for ore grade estimation is extended to 3D CNN for spectrogram representation of drill-core reflectance spectra, in contrast to its more common 2D application. The results showed that it can perform better in 3D than in 2D, and that 3D CNN models are more robust for subtle features that may be present in spectrograms. The broader implications of the results are that ore grade estimation research may benefit from the adoption of 3D CNN models.