Rock type classification based on petrophysical, geochemical, and core imaging data using machine and deep learning techniques

Rock type classification is one of the most crucial steps of geological and geotechnical core logging. In conventional core logging, rock type classification is subjective and time-consuming. This study aims to automate rock type classification using Machine Learning (ML). About 35 m of core samples from five different rock types obtained from an open pit mine were logged using a Multi-Sensor Core Logging (MSCL) system, along with a core scanner that automatically captured geochemical and petrophysical properties of the samples and 360° images of the core circumference. A train/test split strategy (interval split) was introduced, as it produces more realistic predictions than a random shuffle split. The collected logging data were split into train/test subsets based on the core length intervals. For the automated rock type classification, three approaches were implemented. First, different ML algorithms were used to classify rock types based on their petrophysical (P- and S- wave velocities, Leeb hardness) and geochemical properties (collected using a portable X-Ray Fluorescence analyzer (pXRF)). XGBoost outperformed the other models across all rock types. The second approach classified rock types using core images by applying a pre-trained ResNet-50 on ImageNet. Both classical ML and Convolutional Neural Network (CNN) models have higher accuracy for distinct rock samples than transition and interbedding zones. In the third approach, an expert decision procedure was mimicked by concatenating rock properties (first approach) and five features extracted from images (second approach). The concatenation of images and rock properties improved the F1-score of each approach by 10% and 35%, respectively. The core samples had been annotated with a marker in the field, and the effect of removing marked images from the dataset was investigated. The cleaned images improved the rock type prediction by up to 16% (F1-score) using the CNN approach. However, the improvement in the concatenation approach (7%) was not significant enough to justify the labor-intensive cleaning process.