An Efficient and Automated Classification System for Rocks Based on Visually Explainable Deep Learning

Abstract Rock classification is critical in geological research and geoscience applications. Traditional methods rely heavily on manual expertise, which makes them susceptible to human errors due to their reliance on individual skills and experience. Although current machine learning models have mitigated some drawbacks by classifying rock images, their generalization and predictive performance are limited by suboptimal network structures, image data quality, and quantity. These models also require manual feature extraction, increasing training complexity. This article presents an explainable EfficientNet model for eight-class rock classification, pretrained on a novel dataset. Our high-resolution rock specimen images are curated to standardize data, reduce noise, and minimize training perturbations, improving classification precision. To enhance convolutional neural network interpretability and reliability, we further dive into visual interpretation maps generated by various class activation mapping methods. This further demonstrates the model’s generalization capabilities and its ability to capture rock textures, shapes, and colors. This approach not only reinforces the model’s interpretability but also underscores its robustness in identifying key discriminative attributes within rock imagery.