Petrographic image classification of complex carbonate rocks from the Brazilian pre-salt using convolutional neural networks.

Machine learning (ML) algorithms have been widely applied across geosciences for tasks such as data conditioning, resolution enhancement, and image classification. The use of ML enables the analysis of large datasets, the identification of complex patterns, and can save time and reduce costs compared to conventional approaches. Among these techniques, Convolutional Neural Networks (CNNs) have emerged as powerful tools for image classification in various geoscientific applications. In the context of the carbonate reservoirs of the Brazilian Pre-salt, the sedimentological complexity of these deposits, combined with the vast amounts of data produced, drives the need for automated image classification approaches. Although several recent studies have explored ML methods for petrographic image analysis in diverse geological settings, few have focused specifically on the complex carbonates of the Brazilian Pre-salt reservoirs. In this study, we present a fully automated and modular machine learning workflow for petrographic image classification of thin sections from the Aptian Barra Velha Formation, Santos Basin, Brazil. Our approach includes the direct integration of paired plane-polarized light (PPL) and cross-polarized light (XPL) images as raw inputs to deep learning models, allowing for a more comprehensive representation of petrographic features. Additionally, we implement a hierarchical classification scheme, based on facies upscaling, encompassing three levels of classification granularity: a simplified scheme with 5 classes, an intermediate with 9 classes, and a complete scheme with 23 classes, a dimension not systematically explored in previous studies. Our dataset comprises 800 thin sections, corresponding to 1,600 high-resolution scanned images (6,400 dpi), from six wells across three different oilfields, strategically selected to ensure representativeness across distinct structural domains of the reservoir. We evaluated five computational models: EfficientNet, MobileNet v3, RegNet, ResNet, and ShuffleNet v2. The models MobileNet v3 large, RegNet x 800mf, and RegNet y 400mf achieved the highest F1-scores for the simplified (0.795), intermediate (0.768), and complete classifications (0.528), respectively. Notably, the intermediate classification with nine classes offered the best balance between detail and accuracy. This work presents a promising approach for automatic petrographic image pre-classification, favoring efficient database organization in the challenging exploratory settings of the Brazilian Pre-Salt.