Automatic labeling strategy in semi-supervised seismic facies classification by integrating well logs and seismic data

초록

Machine learning algorithms have been widely used for the quantitative seismic interpretation to delineate the distribution of the oil-saturated reservoir. Especially, unsupervised clustering has been more favored than the supervised learning in the early stage of exploration due to the lack of the labeled data such as well logs. However, the unsupervised method performs only grouping the unlabeled data, and requires labeling the facies by human interpreters, so the labeling results can be error-prone and biased depending on the interpreter's ability. To overcome the shortcomings, we propose a new workflow using semi-supervised learning. The fundamental idea is to use labeled and unlabeled data at the same time for the input of the clustering algorithms. In this study, the extended elastic impedance (EEI) well logs and the EEI seismic data are considered as labeled and unlabeled data, respectively. Also, to reduce uncertainty in predicting reservoir properties, we apply the well data augmentation by Monte Carlo simulation based on petro-elastic model. The proposed workflow reduces the intervention of interpreters for labeling the facies, and accurately classify the facies for the seismic volume even when the amount of well log data is small and not sufficient for the supervised learning algorithms.

키워드