Active deep learning for seismic facies classification
Deep Learning, Active Learning, Seismic Interpretation, Image Segmentation
Seismic facies interpretation is a critical aspect of oil and gas exploration, yet it is not practical for human interpreters to thoroughly analyze every part of the data as the volume and resolution of seismic data increases. To address this issue, deep learning-based interpretation methods have gained attention. However, acquiring a sufficiently large and accurately labeled training dataset within project timelines remains challenging. To overcome this obstacle, active learning methods have been proposed. They reduce the number of required training labels by creating an optimized labeled training set from unlabeled data.
In this study, we developed an end-to-end encoding-decoding deep neural network for seismic facies classification and applied an active learning workflow with three distinct query strategies. The research was made using the Parihaka public dataset. Additionally, We introduced a unique bootstrap-based strategy to assess the confidence interval for the active learning curves. Our results showed comparable outcomes to the baseline model could be achieved using less than half of the labeled training dataset, even when employing rudimentary methods, such as random sampling. Notably, uncertainty sampling proved to be the most effective among the query strategies studied, as it has the potential to not only prioritize the most informative images but also identify uninformative ones. These promising findings suggest that incorporating active learning techniques can enhance the practicality and efficiency of deep learning-based seismic interpretation by reducing the reliance on large, labeled training datasets.