Classifying reservoir rock quality from thin sections of core using Convolutional Neural Networks

Abstract: In the past few years, the use of Deep Learning in the petroleum industry has increased, especially for seismic interpretation and signal processing. However its use remains shy in image analysis and carbonates classification. This study aims at applying Deep Learning on reservoir rock images and produce an image classification algorithm. The images used are thin sections of core rocks gathered while drilling exploration wells. The purpose is to classify thin sections according to their level of porosity, their Dunham and Depositional Rock Types (DRT) classification, and to be able to identify the different components they contain. These predictions will assist geologists in their labeling tasks. The algorithm is based on Convolutional Neural Network (CNN). We tested three different CNN with different hyperparamters, optimizers as well as initialization strategies. We compared the results in order to find the best setting to achieve the task. We found that the best performing network was Inception_v3 with random initialization. The best optimizer to use depends on the class we predict on. Our best accuracy on the test set is 76,1% to identify the level of porosity, 68,4% to classify the images according to their Dunham classification and 53,7% for the DRT classification. The network also manages to identify all the components in an image with a F1-score of 50,2%. The main misclassifications are between rocks that are adjacent in the classification. This means that our algorithm struggles to differentiate similar rocks. When we compare our results to experienced geologists with carbonate classification experience, we observe that they make similar mistakes. The results of this study show great potential for the use of Deep Learning in classification of core thin sections.