Application of percentile color intensities of borehole images for automatic fluorite grade assessment

Abstract

This paper describes a practical application of support vector machine (SVM) for ore grading in an underground fluorite mine. It covers all aspects from the inception of the experiments, data collection, input preparation, model description and results. Forty-eight drilling chips samples are collected while drilling six pseudo-horizontal boreholes at depth intervals of half a meter and their chemical composition determined through X-Ray fluorescence; the response of the drill rig is used to accurately define the depth of each sample along the blasthole. Images of the blasthole walls are collected with an optical televiewer with white and ultraviolet (UV) illumination. The color information of the images is characterized by the cumulative distribution of pixel color intensities of red, green and blue, used as inputs. A well-known metaheuristic algorithm is used to calibrate the SVM hyperparameters. Repeated k-fold cross validation is applied to increase the prediction performance due to the small-size of the dataset. An outlier inspection is made resulting in improved performance. The combination of pixel intensities from white and UV light scans leads to the best prediction of fluorite content (average R2 = 0.83 and RMSE = 3.32 %), while intensities from only white light procures the best classification results (average classification accuracies from 0.77 to 1). These metrics support the utility of the proposed methodology for reducing the amount of lab analysis in ore grade control.