A Carbonate Lithology Identification Method Integrating Locally Constrained Data Augmentation and Ensemble Learning

Accurate lithology identification is essential for the exploration and development of carbonate reservoirs. Although machine learning is one of the most effective methods for lithology prediction, the development of transitional lithologies, overlapping logging responses and imbalanced class distributions in carbonate strata often lead to limited identification accuracy when traditional machine learning methods are applied. To address these coupled problems, this study develops a two‐stage, problem‐driven DBOKS–Stacking framework. At the data level, DBO‐stabilized K ‐means partitioning is used to define locally consistent regions before SMOTE interpolation, thereby reducing the risk of cross‐boundary synthetic samples while improving minority‐class representation. These constraints are defined by local statistical consistency in the well‐log feature space rather than by explicitly encoded geological rules. At the model level, Stacking is subsequently introduced to integrate the differentiated decision behaviours of multiple classifiers, addressing the residual prediction instability that cannot be resolved by data balancing alone. The experimental results show that, after DBOKS processing, the overall performance of all models is improved to varying degrees and the identification performance of minority and transitional lithologies is enhanced. After further integration with the Stacking ensemble, the overall accuracy reaches approximately 95%. This study indicates that sample augmentation and multi‐model fusion can improve complex carbonate lithology identification from the perspectives of lithological class imbalance and logging‐response boundary ambiguity, respectively, providing an effective data‐driven method for well‐log‐based lithology classification under complex reservoir conditions.