Ex Situ Hyperspectral Sensing and Machine Learning for Tailings Characterization

Understanding tailings properties at high spatial resolutions is needed for many geotechnical analyses. While tailings properties can be characterized using undisturbed samples and in situ tests, the current methods face limitations regarding adequate spatial resolution and comprehensive property assessment. This study explores the use of hyperspectral sensing and convolutional neural networks for the simultaneous prediction of twelve tailings properties, including percent sand, silt, clay, fines content, solids content, gravimetric moisture content, volumetric moisture content, saturation, void ratio, porosity, total density, and dry density. Tailings from a precious metal mine were used to prepare samples with diverse properties and hyperspectral data were captured. The tailings-hyperspectral dataset was then split into training and testing subsets, a convolutional neural network was optimized and trained, and the model’s performance was assessed using the testing data. The prediction of particle size distribution metrics and moisture metrics have root mean squared errors below 8% with coefficients of determinations above 0.95. Predictions of void ratio, porosity, total density, and dry density have poorer performance than other properties. However, density predictions have lower errors for samples with high saturation. Results shows promise for the rapid characterization of tailings properties using hyperspectral data.