Article: Automated classification of geological core images by extraction region in Algeria: a deep learning approach Journal: International Journal of Computational Vision and Robotics (IJCVR) 2026 Vol.16 No.5 pp.18 - 33 Abstract: Interpretation of geological images is a fundamental process in locating hydrocarbon reservoirs. Accurate information about the source of core samples is essential for efficient and targeted drilling in upcoming campaigns. Traditionally, determining the extraction region of geological core images has depended on the expertise of geologists and geophysicists. To automate this task in the Algerian context, we trained various architectures, including ResNet-50, single-layer CNN, double-layer CNN, triple-layer CNN and a Siamese neural network, using a dataset of real core images collected from multiple drilling campaigns across four regions of Algeria. The triple-layer CNN achieved an impressive accuracy of 100%, alongside perfect precision, recall and F1-score across all classes, while the Siamese neural network exhibited strong confidence in label matching, as evidenced by consistent similarity scores in experiments. This advancement has the potential to significantly expedite the analysis of large volumes of core images, thereby minimising the dependence on human expertise. Furthermore, we are making the adapted models available to the wider community, fostering opportunities for further enhancements. Inderscience Publishers - linking academia, business and industry through research

Title: Automated classification of geological core images by extraction region in Algeria: a deep learning approach

Authors: Rima Boudjadja; Khadidja Belattar; Khaoula Bouzghaia

Addresses: Department of Computer Science, University of Algiers 1, Algiers, Algeria; LaRIA Laboratory, University of Jijel, Jijel, Algeria ' Department of Computer Science, University of Algiers 1, Algiers, Algeria ' Department of Computer Science, University of Algiers 1, Algiers, Algeria

Abstract: Interpretation of geological images is a fundamental process in locating hydrocarbon reservoirs. Accurate information about the source of core samples is essential for efficient and targeted drilling in upcoming campaigns. Traditionally, determining the extraction region of geological core images has depended on the expertise of geologists and geophysicists. To automate this task in the Algerian context, we trained various architectures, including ResNet-50, single-layer CNN, double-layer CNN, triple-layer CNN and a Siamese neural network, using a dataset of real core images collected from multiple drilling campaigns across four regions of Algeria. The triple-layer CNN achieved an impressive accuracy of 100%, alongside perfect precision, recall and F1-score across all classes, while the Siamese neural network exhibited strong confidence in label matching, as evidenced by consistent similarity scores in experiments. This advancement has the potential to significantly expedite the analysis of large volumes of core images, thereby minimising the dependence on human expertise. Furthermore, we are making the adapted models available to the wider community, fostering opportunities for further enhancements.

Keywords: classification; deep learning; convolutional neural network; CNN; Siamese neural network; one shot learning; transfer learning; geological core image; Algeria.

DOI: 10.1504/IJCVR.2026.153921

International Journal of Computational Vision and Robotics, 2026 Vol.16 No.5, pp.18 - 33

Received: 11 May 2024
Accepted: 18 Oct 2024
Published online: 08 Jun 2026 *

Title: Automated classification of geological core images by extraction region in Algeria: a deep learning approach

Keep up-to-date