Rock-type classification: A (critical) machine-learning perspective

We investigate machine-learning techniques for rock-type classification. A throughout literature review (considering the machine-learning technique, number of classes, rock types, and image types) presents a diversity of datasets employed and a wide range of classification results as well as multiple problem formulations. Throughout the discussion of the literature, we highlight some common machine-learning pitfalls and criticize the decisions taken by some authors on the problem formulation. We present an experimental contribution by evaluating the classification of seven types of rocks found in carbonate reservoirs along with state-of-the-art Convolutional Neural Networks (CNNs) architectures available through a well-known open-source library. For this experimentation, we detail the preparation of the dataset of drill core plugs (DCPs), the experimental setup itself, and the obtained results considering the normalized accuracy and the traditional accuracy as metrics. We performed the manual background segmentation of the employed dataset of DCPs; so the results reported are not influenced by the background of the images. We evaluate top-1, top-2, and top-3 performance for the problem. We apply fusion of multiple CNNs for richer classification decisions. We also contribute by presenting the manual classification — human labeling by looking at the image on the computer screen — of the same seven-class dataset, performed by six non-geologist volunteers. Finally, we present a conclusion for the results obtained with our experiments and share valuable advice for researchers applying machine learning to rock classification.