Scholarlink

<p>The reliable and accurate identification of faults in seismic images serves as the foundation for seismic interpretation. This paper presents an interpretable tree-based ensemble model for identifying seismic faults using the gradient boosting algorithm and Bayesian optimization. First, seismic attributes that are sensitive to discontinuity features were selected, including amplitude-based, similarity, and semblance attributes. Then, to address the issue of information redundancy caused by multiple seismic attributes, a feature optimization algorithm was applied to reduce the dimensionality of seismic attribute data. The effectiveness of various gradient boosting algorithm models in identifying seismic faults was subsequently evaluated. These included gradient boosting decision trees (GBDT), extreme gradient boosting (XGBoost), CatBoost, and light gradient boosting machine (LightGBM). The integration of SHapley Additive exPlanation (SHAP) to analyze the impact and dependency between features and prediction results enhances the interpretability of the model, thereby providing reliable support for the prediction results. The efficacy of the proposed method was subsequently evaluated through its application to the test datasets. The findings of the study indicated that the fault prediction accuracy of the comparative ML models follows a clear hierarchy: CatBoost/LightGBM&gt;XGBoost/GBDT&gt;MLP&gt;SVM. The CatBoost and LightGBM algorithms, which are based on Bayesian optimization, achieved a test accuracy of 0.93, while XGBoost and GBDT achieved 0.92. All tree-based ensemble models exhibited superior performance in comparison to traditional non-ensemble models, including multi-layer perceptron (MLP, 0.85) and support vector machine (SVM, 0.76). Specifically, BO-CatBoost demonstrated superior performance across the key indicators (Precision, F1 score and AUC), confirming the value of Bayesian optimization in enhancing tree-based model performance. The study demonstrates the interpretable tree-based ensemble model’s superior precision and efficiency in seismic fault detection.</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

Seismic Fault Detection Using Interpretable Tree-Based Ensemble Models with Bayesian Optimization

  • 期刊论文

  • 发表日期：

    2025年11月8日

查看全文
  • Yang Li,
  • Suping Peng,
  • Xiaoqin Cui,
  • Peng Lin,
  • Dong Li,
  • Tao He,
  • Kunheng Li,
  • Henggao Geng,
  • Hongpeng Zhao

摘要

The reliable and accurate identification of faults in seismic images serves as the foundation for seismic interpretation. This paper presents an interpretable tree-based ensemble model for identifying seismic faults using the gradient boosting algorithm and Bayesian optimization. First, seismic attributes that are sensitive to discontinuity features were selected, including amplitude-based, similarity, and semblance attributes. Then, to address the issue of information redundancy caused by multiple seismic attributes, a feature optimization algorithm was applied to reduce the dimensionality of seismic attribute data. The effectiveness of various gradient boosting algorithm models in identifying seismic faults was subsequently evaluated. These included gradient boosting decision trees (GBDT), extreme gradient boosting (XGBoost), CatBoost, and light gradient boosting machine (LightGBM). The integration of SHapley Additive exPlanation (SHAP) to analyze the impact and dependency between features and prediction results enhances the interpretability of the model, thereby providing reliable support for the prediction results. The efficacy of the proposed method was subsequently evaluated through its application to the test datasets. The findings of the study indicated that the fault prediction accuracy of the comparative ML models follows a clear hierarchy: CatBoost/LightGBM>XGBoost/GBDT>MLP>SVM. The CatBoost and LightGBM algorithms, which are based on Bayesian optimization, achieved a test accuracy of 0.93, while XGBoost and GBDT achieved 0.92. All tree-based ensemble models exhibited superior performance in comparison to traditional non-ensemble models, including multi-layer perceptron (MLP, 0.85) and support vector machine (SVM, 0.76). Specifically, BO-CatBoost demonstrated superior performance across the key indicators (Precision, F1 score and AUC), confirming the value of Bayesian optimization in enhancing tree-based model performance. The study demonstrates the interpretable tree-based ensemble model’s superior precision and efficiency in seismic fault detection.