Feature-Depth Smoothness Based Semi-Supervised Weighted Extreme Learning Machine for lithology identification

Abstract

Lithology identification is a fundamental task in the field of hydrocarbon exploration, which is especially essential to understand the subsurface geology. These years have witnessed the development of intelligent lithology identification using machine learning algorithms, but most of them are investigated by using supervised methods. However, considering that getting the core and cuttings samples is quite costly, and one cannot capture all the samples along the well, only partial well logging data could be labelled, thus motivating us to study semi-supervised learning in this work. In this paper, we propose a semi-supervised learning method named Feature-Depth Smoothness Based Semi-Supervised Weighted Extreme Learning Machine (FD-S2WELM) to solve the lithology identification problem with scarce labels. The contributions are threefold: (i) Considering the unbalanced distribution of subsurface lithologies, the weighting mechanism is introduced to enhance the identification of the minority classes. (ii) Both the smoothnesses in feature space and depth are formulated and introduced to construct the graph Laplacians, and thus increasing the safety of the introduction of smoothness regularization. (iii) A large number of experiments are conducted to exhibit the classification performance over different parameters and labels, and thus demonstrating the superiority in safety of our method. As demonstrated in the experiments, the proposed FD-S2WELM performs the best under most experimental settings. Notably, unlike the existing semi-supervised methods which may suffer from unsafe issues (i.e., semi-supervision is worse than supervision), the worst-case accuracies of FD-S2WELM are never lower than any supervised methods, thus presenting high safety or robustness against uncertainty.