Proof of Concept for Dunham Classification of Carbonate Thin Sections Using Vision Transformers

The classification of carbonate rocks using petrographic thin sections is a fundamental task in sedimentology and is traditionally performed by geologists using schemes such as Dunham’s depositional-texture classification. However, this approach is time-consuming and may be prone to inter-observer variability, as reported in prior studies. This paper presents a proof of concept for using a Vision Transformer (ViT) to classify carbonate thin-section images into Dunham classes. The dataset from Lokier and Al Junaibi (2016) is used for fine-tuning and consists of 14 carbonate samples imaged at multiple magnifications and polarization states and labeled through expert consensus. A pre-trained ViT model (vit-base-patch16-224-in21k) is fine-tuned using open-source tools, with an emphasis on reproducibility through transparent methodology and publicly available code. The fine-tuned model achieves high accuracy on a held-out test split derived from the same dataset; however, this performance likely reflects overfitting due to the limited size of the dataset. Evaluation on an independent validation dataset sourced from the CarbonateWorld website yields a lower accuracy of 67%, consistent with known ambiguities in carbonate texture classification. Notably, when considering the top-2 predicted classes, validation accuracy improves to 85.71%, indicating that secondary prediction probabilities capture these ambiguities. Overall, the results highlight both the potential of ViT-based models for carbonate classification and the importance of open, expert-labeled datasets. Rather than proposing a state-of-the-art classifier, this study demonstrates a reproducible and extensible workflow. By providing a clear methodology and code snippets, it enables readers to fine-tune ViT models for their own use cases.