Feature-Weighted and Dynamically Constructed Random Forest for Lithology Identification in Intelligent Oilfield Exploration

Automatic lithology classification from well-logging data is a critical task in intelligent oilfield exploration, enabling efficient reservoir evaluation and geological modeling. To address the limitations of traditional methods in capturing key features and adapting to complex geological structures, this paper proposes a Feature-Weighted and Dynamically Constructed Random Forest (FW-DC-RF) for lithology identification. The proposed model first calculates feature importance scores from an initial random forest and applies them to construct a feature-weighted input matrix, emphasizing the contribution of critical logging parameters. Then, a boosting-inspired dynamic training strategy is employed to iteratively increase the weights of misclassified samples, guiding the model to focus on hard-to-classify or minority lithology classes. Experiments on a real-world logging dataset (LB07A) demonstrate that the proposed method outperforms the standard random forest in accuracy, precision and F1-score, achieving an average accuracy improvement of over $3.5 \%$. The method features a clear structure and straightforward implementation, making it interpretable and practical for deployment in intelligent lithology classification under complex geological conditions.