Evaluating the Prediction Performance of Random Forest in Classification of Carbonate Lithology

Abstract

Accurate lithology prediction in carbonate reservoirs is essential for hydrocarbon exploration but remains challenging due to their complex heterogeneity. Traditional methods (e.g., seismic and well-log analysis) often fail to capture subtle lithological variations, while machine learning approaches such as Random Forest (RF) remain underexplored for carbonates. Previous research has not sufficiently compared Random Forest with advanced models such as XGBoost and deep learning approaches, nor provided detailed feature importance analyses specific to carbonate lithology classification. This study employs a dataset comprising 4,624 samples characterized by ten petrophysical properties to evaluate the classification performance of RF. Our optimized RF framework demonstrates superior accuracy while reducing dependence on costly core sampling, thereby improving the precision of carbonate reservoir models.