Scholarlink

<p>Lithology identification is a fundamental task in seismic reservoir characterization. However, existing studies have primarily focused on optimizing single algorithms, with limited systematic comparisons of ensemble models and hyperparameter optimization strategies. To address this issue, this study, based on well seismic data from the North Sea F3 block, integrates recursive feature elimination with cross-validation (RFECV), the Near-Miss&#xa0;Synthetic Minority Oversampling Technique (NM-SMOTE) sampling strategy, and four mainstream hyperparameter optimization methods to evaluate the performance of random forest, XGBoost, LightGBM, CatBoost, and stacking ensemble method. NM-SMOTE (Near-Miss SMOTE) effectively alleviates the class imbalance problem by synthesizing minority sandstone samples and retaining key mudstone samples&#xa0;that are closest to the minority class (while reducing the majority class size), thereby improving the reliability of minority class recognition. Fivefold cross-validation was employed, using well log lithology interpretation as ground truth for validation. The results indicate that Optuna achieves the best balance between efficiency and accuracy, outperforming Bayesian optimization and grid search in terms of test accuracy, training time, and model stability. CatBoost achieves the highest prediction accuracy (area under the receiver operating characteristic curve, AUC = 0.91), demonstrating clear sandstone–mudstone boundaries and superior continuity in predictions. These findings provide a reliable basis and methodological support for the selection and optimization of intelligent lithology identification models under complex geological conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimization and evaluation of ensemble learning models for intelligent lithology identification based on seismic data

  • 期刊论文

  • 发表日期：

    2025年12月10日

查看全文
  • Wang Jingyi,
  • Jiang Li,
  • Feng Zhibing,
  • Huang Xiao,
  • Yao Zhenan,
  • Zhang Bocheng

摘要

Lithology identification is a fundamental task in seismic reservoir characterization. However, existing studies have primarily focused on optimizing single algorithms, with limited systematic comparisons of ensemble models and hyperparameter optimization strategies. To address this issue, this study, based on well seismic data from the North Sea F3 block, integrates recursive feature elimination with cross-validation (RFECV), the Near-Miss Synthetic Minority Oversampling Technique (NM-SMOTE) sampling strategy, and four mainstream hyperparameter optimization methods to evaluate the performance of random forest, XGBoost, LightGBM, CatBoost, and stacking ensemble method. NM-SMOTE (Near-Miss SMOTE) effectively alleviates the class imbalance problem by synthesizing minority sandstone samples and retaining key mudstone samples that are closest to the minority class (while reducing the majority class size), thereby improving the reliability of minority class recognition. Fivefold cross-validation was employed, using well log lithology interpretation as ground truth for validation. The results indicate that Optuna achieves the best balance between efficiency and accuracy, outperforming Bayesian optimization and grid search in terms of test accuracy, training time, and model stability. CatBoost achieves the highest prediction accuracy (area under the receiver operating characteristic curve, AUC = 0.91), demonstrating clear sandstone–mudstone boundaries and superior continuity in predictions. These findings provide a reliable basis and methodological support for the selection and optimization of intelligent lithology identification models under complex geological conditions.