Data augmentation using CycleGAN for overcoming the imbalance problem in petrophysical facies classification

Abstract

The petrophysical facies classification in the field of hydrocarbon exploration is one of the important tasks for reservoir characterization. To predictthe facies of the seismic area, deep learning has recently been applied. However, when applying machine learning(ML)to the facies classification, there is a problem that the data available for training are very limited.Whenusing training data acquired under such limited conditions, such as well log data, there can be a severe imbalance in the number of training samples forthe faciesbecause the amount of data acquired in the hydrocarbon area of interest is relatively less than that acquired in the nonhydrocarbon area. Thus, the facies classification results often show weighted predictions of a specific facies due to the imbalance issueof training data.In this study, to solve thelimitation and the imbalance problemsof training data, the data augmentation techniquebased on CycleGAN (Cycle-consistent Generative Adversarial Networks)was conducted. Using CycleGAN, the training data for the class with less data can be augmented using the class with more data. By comparing the results of the faciesclassificationwith and without the augmentation of the training data using Cycle GAN,we have demonstrated that the more accurate classification model is trained when the CycleGAN is applied. Therefore, the data augmentation schemedeveloped in this study will be very useful for thefacies classification in an environment where training datais very limited.