LithoVision: a novel approach for carbonate rock classification in the Brazilian pre-salt

Abstract Classifying lithotypes from a reservoir is crucial for assessing drilling, production, and operational risks. Traditionally, accurately identifying the rock lithologies has been a manual, time-consuming, and error-prone process. Recent advancements in deep learning-based and computer vision have shown promising results in classifying rock lithology. However, many state-of-the-art techniques struggle when working with limited, imbalanced, and fine-grained rock datasets, revealing room for improvement. To address these limitations, this paper proposes an innovative strategy that integrates the Vision Transformer and the Central Difference Convolution architectures to classify the major lithology classes in carbonate rock images. The proposed method combines an Encoder Transformer with a Convolutional Neural Network to learn invariant representations of carbonate textures, eliminating manual feature extraction and performing effectively for small and large datasets. We conducted extensive experimental evaluations on two real-world datasets, demonstrating that: (1) our approach is competitive on fine-grained rock datasets; (2) it surpasses most baseline methods in most evaluation scenarios. Specifically, our proposal outperforms the second-best method (ViT) by 8% in the whole-image classification, and exceeds the best CNN model (VGG) by at least 24% in patches classification, considering the balanced accuracy metric.