Facies classification is a way to gain an in-depth understanding of the geology by revealing details of underlying geologic features. It is critical to define lithology of interest to build a better understanding of depositional environment across the survey area. Core samples and well log data of reservoir rock are the primary sources for facies information, but they come at very high cost. Furthermore, the conventional methods of manually assigning lithofacies in seismic area is time consuming and can be impractical due to immense volume of seismic data. This has motivated machine learning approach for time and cost-efficient classification of seismic facies from seismic inverted data. However, labeled data (well data) are very expensive to obtain and are limited by various constraints, whereas there is a plethora of unlabeled data (seismic data). Training of deep learning (DL) models with relatively sparse labeled datasets for facies classification is a critical issue these days. To this end, the design of proper regularization techniques plays a central role. In this study, we address the problem scarcity of labeled data for improving generalization performances of DL models with semi-supervised learning. We introduce a novel model for leveraging unlabeled data to improve generalization performances for rock facies classification. We used an additional reconstruction module with a conventional DL model to regularize and encourage the feature extraction of the input data. With the addition of reconstruction loss to the total loss, the model leverages unlabeled data for the task of feature learning. Using our proposed model, we show that the model can accurately classify facies using a limited and class imbalanced of training dataset while outperforming conventional DL models.

This content is only available via PDF.
You can access this article if you purchase or spend a download.