Synergizing AI with Geology: Exploring VisionTransformers for Rock Classification

Abstract Automated rock identification is a critical domain within geosciences, offering significant advancements over traditional manual methods. This study explores the benefits of automation in rock identification, highlighting improvements in speed, efficiency, consistency, and accuracy. Such advancements facilitate large-scale geological surveys and allow for the seamless integration of various geoscientific data, yielding novel insights into Earth’s geology and resources. The research presents a robust model that achieves an impressive accuracy rate of 88.00%, outperforming conventional Convolutional Neural Networks (CNN) and Transfer Learning approaches. A comprehensive analysis of performance metrics, confusion matrices, and ROC-AUC curves demonstrates the model’s robustness and precision in classifying different rock types. Despite its success, the study acknowledges the challenges in accurately classifying Basalt and suggests potential solutions such as data augmentation and incorporating domain-specific knowledge. The conclusion emphasizes the need for future research to address class imbalances and to adopt more advanced techniques for feature extraction and classification, aiming to further refine the model’s accuracy and applicability in automated rock identification