Application of Dimensionality Reduction Algorithm for Lithofacies Classification in a Heteroscedastic Well Log Dataset

Predicting the lithofacies using the well log responses is an important step in exploration geophysics. Using machine learning (ML) techniques, the task becomes more time-efficient and robust than manual interpretation for a large dataset. The subsurface environment of the Krishna-Godavari (KG) basin is mainly occupied by shale areas with some trapped water and hydrocarbon zones. The target lithofacies classes are divided into three categories based on their sand concentration, further divided into brine, oil, and gas. With numerous classes and much-skewed distribution, the traditional ML classification algorithm fails to predict all the classes. Scaling cut is a supervised dimensionality reduction algorithm suitable for this kind of heteroscedastic dataset. This work applies scaling cut to find the optimum projection axes with maximum class separability in the multidimensional feature space of well-log attributes. Then ML classification algorithm is applied in a reduced feature dimension to predict the lithofacies classes. The approach has better accuracy than traditional ML approaches in predicting sparsely distributed hydrocarbon classes on KG basin dataset.