Automated classification of drill core photographs: A complete workflow for weakly supervised training

Recent studies have shown the interest of automating the classification of rock images using deep learning architectures. The biggest issue for practitioners when applying these methods to real-world data sets generated during mineral exploration is the long time required to create and label a data set. This study proposes a complete workflow to label and classify drill core photographs with minimal time required for labeling through five successive steps: i) using exploration drill core photographs, rock cores are separated from wooden trays using morphological operators; ii) feature descriptors are then extracted from rock images using color histograms for colorimetric information and Gabor filters for texture information; iii) features extractors then serve as input data for self-organizing maps (SOM) for generating clusters that can be partially labeled by geologists for generating a labeled data set with limited efforts, generating a data set made of labeled and unlabeled images; iv) the partially labeled data set can then be used to train either fully supervised or semi-supervised deep learning architectures for generating classifications; v) the classification model obtained can then be re-used on unseen data to automate logging process. This study presents this workflow separately for two different geological domains, namely a data set of sedimentary rocks classified according to the intensity of bleaching features and a data set of crystalline basement rocks classified by lithological domains. Software code and data set are made publicly available. • Development of a complete workflow to transform drill-core photographs datasets (directly obtained from mineral exploration) into a labeled dataset usable for deep-learning tasks. • Evaluation of the possibility of inferring lithological (for the basement subset) and alteration (for the basin subset) classes based on a self-labeling strategy. • Evaluation of the potential to improve classification accuracy by using unlabeled samples in a semi-supervised way.