Data-Driven Algorithms for Image-Based Rock Classification and Formation Evaluation in Formations With Rapid Spatial Variation in Rock Fabric

Supervised learning algorithms can be employed for the automation of time-intensive tasks, such as image-based rock classification. However, labeled data are not always available. Alternatively, unsupervised learning algorithms, which do not require labeled data, can be employed. Using either of these methods depends on the evaluated formations and the available training/input data sets. Therefore, further investigation is needed to compare the performance of both approaches. The objectives of this paper are (a) to train two supervised learning models for image-based rock classification employing image-based features from computerized tomography (CT) scan images and core photos, (b) to conduct image-based rock classification using the trained model, (c) to compare the results obtained using supervised learning models against an unsupervised learning-based workflow for rock classification, and (d) to derive class-based petrophysical models for improved estimation of petrophysical properties First, we removed non-formation visual elements from the core image data, such as induced fractures, the core barrel, and the seal peel tag on core photos. Then, we computed image-based features such as grayscale, color, and textural features from core image data and conducted feature selection. Then, we employed the extracted features for model training. Finally, we used the trained model to conduct rock classification and compared the obtained rock classes against the results obtained from an unsupervised image-based rock classification workflow. This workflow uses image-based rock fabric features coupled with a physics-based cost function for the optimization of rock classes. We applied the workflow to one well intersecting three formations with rapid spatial variation in rock fabric. We used 60% of the data to train a random forest and a support vector machines classifier using a 5-fold cross-validation approach. The remaining 40% of the data was used to test the accuracy of the supervised models. We established a base case of unsupervised learning rock classification and four different cases of supervised learning rock classification. The highest accuracy obtained for supervised rock classification was 97.4%. The accuracy obtained in the unsupervised learning rock classification approach was 82.7% when compared against expert-derived lithofacies. Class-based permeability estimates decreased the mean relative error by 34% and 35% when compared with formation-based permeability estimates, for the supervised and unsupervised approaches, respectively. The highest accuracies for the supervised and unsupervised models were obtained when integrating features from CT-scan images and core photos, highlighting the importance of feature selection for machine-learning workflows. A comparison of the two approaches for rock classification showed higher accuracy obtained from the supervised learning approach. However, the unsupervised method provided reasonable accuracy as well as a more general and faster approach for rock classification and enhanced formation evaluation.