Semi-supervised learning for lithology identification using Laplacian support vector machine

Lithology identification is a fundamental task in well log interpretation. Considering the presence of substantial unlabeled data in the field of petroleum exploration, this paper investigates the semi-supervised learning method for lithology identification, and proposes a semi-supervised lithology identification workflow. In the workflow, the Laplacian support vector machine is employed to achieve semi-supervised learning. The feature similarity and depth similarity are introduced to reveal the data distribution characteristic, thus enabling us to elevate the classification performance of the Laplacian support vector machine dealing with the issue of lacking labels. Considering the well logging dataset without label, K-means clustering is used to select the k well logging samples that should be labeled by experts, based on which Laplacian support vector machine algorithm can be then conducted. The proposed method is compared with the supervised method (e.g., support vector machine) that excludes the unlabeled data and the semi-supervised method without considering the depth similarity. The comparison experiments are conducted on four datasets that are collected from the Jiyang Depression, Bohai Bay Basin, China. The experimental results show that the utilization of unlabeled data can improve identification performance, especially that of minority lithology classes. It is also verified that the information provided by feature similarity and depth similarity are helpful for lithology identification.