Classification of igneous rocks from petrographic thin section images using convolutional neural network

Rock classification from petrographic thin section analysis often requires expertise in mineralogy. This study developed a deep learning approach based on a convolutional neural network (CNN) to classify six igneous rock types from their thin section images. Petrographic image dataset with various image conditions was prepared and processed to train and evaluate the network model. The results from two different test methods demonstrated that the classification accuracy was higher when the classification scores of partitioned image patches were summed for an original image (Test A method) than when those of each partitioned image patch were individually predicted (Test B method). Nevertheless, both methods resulted in higher than 90% accuracy, proving that partitioned image-based classification could be suitable for petrographic images with various conditions. The features identified by the ResNet152 model were qualitatively evaluated by applying gradient-weighted class activation mapping (Grad-CAM) to the last convolutional layer. The correctly classified images showed well-perceived mineral grains and the associated matrix as visualized by Grad-CAM. It implied that CNN-based models could successfully identify morphological characteristics within an image similar to the human-based approach, leading to a reliable and explainable method for rock classification.