Benchmarking deep CNN architectures for micro-CT rock image classification

This study presents a benchmark analysis of deep convolutional neural network (CNN) architectures for rock type classification using high-resolution two-dimensional grayscale slices extracted from 3D micro-computed tomography (µCT) datasets. The main contribution of this work is the development of a controlled and systematic benchmarking framework for CNN architectures under standardized µCT imaging conditions, rather than expanding the number of lithological classes or facies. A dataset comprising 52,706 two-dimensional grayscale µCT slices derived from 11 sandstone samples, 16 limestone samples, and eight shale plugs extracted from a single core was used as a three-class geological classification problem. Seventeen pretrained CNN architectures were evaluated under a standardized workflow incorporating data augmentation, area-like image downsampling, and specimen-level grouped three-fold cross-validation to reduce data leakage and account for the physical sampling structure. Model performance was assessed using accuracy, loss, weighted precision and weighted recall, weighted F1-score, mean per-class accuracy (MPCA), confusion matrices, specimen-level majority-vote metrics, and learning curves. Several architectures, including DenseNet169, DenseNet121, DenseNet201, and InceptionV3, achieved reliable classification performance across all metrics with stable convergence behavior. Sandstone and limestone samples were the most frequently confused classes, whereas shale samples were classified more consistently within the available dataset. Comparison with previous studies highlights recurring challenges such as intra-class variability, overfitting in small datasets, and the influence of image modality on model performance. Overall, the results provide a reproducible Earth science informatics framework for evaluating deep learning models in micro-CT-based rock image classification and digital rock analysis.