In a comprehensive comparative study evaluating deep learning models for petrographic analysis, convolutional neural networks classified rock thin-section images with single-split test accuracies ranging from 97.72 percent to 98.86 percent. Under this image-level splitting strategy, the performance gap between the highest-scoring and lowest-scoring architectures corresponded to a difference of only three correctly classified test images. Furthermore, pairwise comparisons between the tested models revealed no statistically significant differences in performance once corrections for multiple testing were applied.
The investigation addressed the classification of petrographic thin sections, an essential geological task that is conventionally time-consuming and highly dependent on specialized human expertise. Over the past decade, rock identification utilizing deep learning methodologies has seen significant advances across geological research. Applying computational techniques to the analysis of rock thin-section images offers a practical means to support petrographic workflows by providing fast and reproducible classifications across diverse samples.
To examine how various network designs perform when categorizing samples into the three major rock classes of igneous, metamorphic, and sedimentary rocks, the authors evaluated eight distinct convolutional neural network architectures. The selection comprised both traditional and modern models, specifically ConvNeXt-Tiny, DenseNet121, EfficientNet B0, Inception V3, MobileNet V3 Large, ResNet50, VGG16, and VGG19. This selection enabled a direct comparison of structural paradigms on identical rock thin-section classification tasks.
The researchers tested the models using two rigorous evaluation schemes: single-split testing and k-fold cross-validation. In cross-validation experiments, the modern ConvNeXt-Tiny architecture attained the highest numerical mean accuracy rates among all evaluated networks, achieving 99.01 percent in five-fold cross-validation and 99.20 percent in ten-fold cross-validation. In comparison, traditional architectures, namely VGG16 and VGG19, generally produced lower numerical mean accuracies, although these numerical variations did not reach statistical significance following multiple testing corrections.
Alongside standard performance metrics, the study conducted a detailed assessment of computational cost to determine practical feasibility for geological applications. The computational demands of the eight convolutional neural networks were characterized using four distinct operational criteria: total parameter counts, floating-point operations measured in GFLOPs, graphics processing unit throughput, and inference latency.
To move beyond black-box predictions and illuminate how the convolutional neural networks identified petrographic features, the researchers integrated Explainable Artificial Intelligence methods into their workflow. The decision mechanisms of the trained models were interpreted using two visual explanation techniques: Grad-CAM and Score-CAM. Applying these methods allowed the authors to analyze how the networks interpreted thin-section imagery when distinguishing between igneous, metamorphic, and sedimentary rock types.
By combining accuracy metrics, computational cost profiling, and decision mechanism interpretations, the study presents an automated approach to petrographic thin-section classification that is both accurate and explainable. The findings demonstrate that while modern architectures such as ConvNeXt-Tiny deliver top numerical accuracy rates up to 99.20 percent, multiple traditional and modern convolutional neural network architectures achieve high, statistically comparable performance when identifying the three primary rock classes.
Subject of Research: Deep learning classification and explainability for petrographic rock thin sections
Article Title: Classification of rock thin-section images into igneous, metamorphic and sedimentary rocks: a comparison of CNN architectures and XAI methods
Article References: Tunç, G., Özsaraç, Ş., Taşlı, S., & Yıldırım, G. (2026). Classification of rock thin-section images into igneous, metamorphic and sedimentary rocks: a comparison of CNN architectures and XAI methods. Earth Science Informatics, 19(12), Article 251. https://doi.org/10.1007/s12145-026-02307-3
Image Credits: AI Generated
DOI: 10.1007/s12145-026-02307-3
Keywords: convolutional neural networks, petrographic thin sections, rock classification, igneous rocks, metamorphic rocks, sedimentary rocks, ConvNeXt-Tiny, DenseNet121, EfficientNet B0, Inception V3, MobileNet V3 Large, ResNet50, VGG16, VGG19, Grad-CAM, Score-CAM, explainable artificial intelligence, cross-validation
News Source: Denise Maddox. (November 17, 2026). ConvNeXt-Tiny Achieves Top Accuracy in Classifying Rock Thin-Section Image Types. Scienmag.



