Vision transformers as SOTA models for lithological classification of Brazilian pre-salt rocks

This paper addresses the classification of pre-salt rock lithology using vision transformers, focusing on evaluating the learning and generalization capabilities of state-of-the-art pre-trained models under conditions of limited data and significant class imbalance. Our research investigates the performance of ten ViT architectures, fine-tuned on a private dataset of petrographic thin section images of carbonate rocks from two Brazilian oil wells. We also explore the impact of two different patch sizes (200 × 200 and 150 × 150 pixels) on model performance. The experimental results demonstrate that DaViT and MViTv2 achieve the highest accuracy and consistency, with DaViT surpassing 95% accuracy with the smaller patch size. In contrast, models like EfficientViT and FlexiViT exhibit lower performance and higher variability, indicating a need for further optimization. This study highlights the potential of vision transformers for geological applications and establishes new benchmarks for automated lithological classification, providing valuable insights for future research in enhancing the efficiency and accuracy of lithological studies.