Automatic Lithology Classification of Whole Core Images Using Multi-Input Convolutional Neural Networks: Integrating Visible Light and Ultraviolet Photography

Abstract Whole core photography is an essential step in core analysis, offering complementary insights that improve accuracy of lithological assessments. Cores are photographed to document their features accurately, preserve a visual record for future analysis, and facilitate better identification of lithological units. Visible (or white) light photography provides detailed visual information about rock physical characteristics, while ultraviolet- (UV-)light images are commonly used due to the fluorescence properties of certain minerals, enhancing mineral detection and identification. Both approaches improve the accuracy of lithological assessments and inform subsurface management decisions. With the rise of artificial intelligence and the demand for precise and automatic rapid predictions, machine learning techniques, particularly convolutional neural network (CNN) architectures, have gained prominence in lithology identification research. This study implemented multi-input CNNs to automatically predict lithology from whole core images. To improve image classification accuracy, we combined UV- and white-light images as input to the CNN, allowing the network’s filters to learn richer features automatically. We used 176-m core data from two formations in the Middle East. Data augmentation techniques were used to create 6000 images. The dataset was randomly divided into three parts for training, validation, and testing the networks. We selected the ResNeXt-50 architecture for its superior efficiency in classifying three lithologies: sandstone, loose sand, and limestone. This model was compared to the EfficientNet architecture. The network parameters were initialized using transfer learning. We optimized the network hyperparameters including learning rate, batch size, and optimizer, achieving 99% accuracy in predicting unseen data. This study establishes an accurate and rapid procedure for automatic lithology classification, outputting a lithology column.