Explainable Artificial Intelligence Using Convolutional Neural Networks for Sedimentary Structure Classification

Convolutional Neural Networks (CNNs) offer powerful image classification capabilities in geology, where visual interpretation of sedimentary structures is essential for reconstructing depositional environments. However, their “black-box” nature can limit trust and slow adoption in geoscientific workflows. This study presents a CNN-based framework for classifying eleven physical and biogenic sedimentary structures from core images, while treating explainability as a core objective. Three architectures (EfficientNet-B2, MobileNet-V3, and ResNet-50) were trained and evaluated on a geologically diverse annotated dataset. EfficientNet-B2 achieved the highest overall accuracy (97.7%), followed by MobileNet-V3 (96.8%) and ResNet-50 (95.8%). Blind-test evaluation on 401 unseen images showed that the CNN models retained substantial predictive capability under domain-shift conditions, with accuracies of 79.05% for EfficientNet-B2, 78.8% for MobileNet-V3, and 78.3% for ResNet-50. These results indicate good overall generalization, while the remaining errors were mainly associated with geologically similar sedimentary structures. Explainability was assessed using Grad-CAM++ and, crucially, was evaluated quantitatively using (i) deletion-based faithfulness, measuring the drop in target-class probability after masking the most salient regions, and (ii) stability under controlled perturbations (brightness changes, noise, blur, and small spatial shifts), measuring similarity between explanation maps. Across models, Grad-CAM++ consistently emphasized geologically meaningful features (e.g., pebble boundaries, inclined laminae, and biogenic textures) while de-emphasizing non-diagnostic artifacts. ResNet-50 produced the most faithful and stable explanations, showing the steepest probability decay under deletion and the highest robustness of saliency patterns under perturbations, despite slightly lower classification accuracy. Overall, these results demonstrate that high classification accuracy does not necessarily imply high explainability quality, and that explanation reliability metrics provide a practical route to more transparent and trustworthy sedimentary structure classification for real-world core analysis.