Facies classification using semi-supervised deep learning with pseudo-labeling strategy

Abstract

Quantitative facies classification is the key to linking seismic data to lithology to evaluate important reservoir properties. During the past several years, the size of seismic volumes has piled up to the extent that it is challenging for experts to examine every seismic volume to classify the facies. This has motivated machine learning approach for predicting seismic facies in an efficient way. However, labeled data (well data) is limited by various constraints and is very expensive to obtain, whereas, there is a plethora of unlabeled data (seismic data). Geophysicists are tasked to interpret enormous amount of unclassified data on the basis of sparse amount of labeled data. In this study, we have adopted Semi-Supervised Learning using pseudo-labeling to facies analysis in order to overcome the scarcity of labeled data by leveraging unlabeled data. With each step, a small amount of data is classified starting from the vicinity of the well and gradually moving away from the well. The classified data (called 'pseudo label data') are added to the label data used in retraining the classifier, adding the diversity to the classifier that accounts for lateral change in lithology while moving away from well. Following our proposed workflow, we have shown that the accuracy of a trained classifier on limited amount of labeled data can be enhanced considerably by combining a small number of labeled well data with a large pool of inverted seismic data using pseudo-labeling technique. Furthermore, with results of applying the proposed workflow to field data, outperforming conventional methods, we have achieved an accuracy of 99.69% and loss as low as 0.001 for both training and validation, moreover, classification task is carried out with error as low as 0.004.