Automated Rock Quality Designation Using Convolutional Neural Networks

Abstract

Mineral and hydrocarbon exploration relies heavily on geological and geotechnical information extracted from drill cores. Traditional drill-core characterization is based purely on the subjective expertise of a geologist. New technologies can provide automatic mineral analysis and high-resolution drill core images in a non-destructive manner. However, automated rock mass characterization presents a significant challenge due to its lack of generalization and robustness. To date, the automated estimation of rock quality designation (RQD), a key parameter for rock mass classification, is based mostly on digital image processing techniques with significant user biases. Alternatively, we propose using computer vision and machine learning-based algorithms for drill core characterization using drill core images to determine the RQD. A convolutional neural network (CNN) is used to detect and classify intact and non-intact cores, and to filter out empty tray areas and non-rock objects present in the core trays. The model calculates the length of the detected intact cores and estimates the RQD. We train the CNN model with thousands of sandstone core images from different drill holes in South Australia. The proposed method is tested on 540 sandstone core rows and 90 limestone core rows (~ 1 m each), which produces average error rates of 2.58% and 3.17%, respectively.