Scholarlink

<p>This study aims to enhance the domain of slope stability assessments by integrating unsaturated soil mechanics into machine learning (ML) methodologies. It addresses both regression and classification problems to predict the factor of safety (FOS) without the need to determine the critical slip surface iteratively. Orthogonal Latin hypercube sampling is employed to generate a comprehensive dataset of 16,219 data points. This synthetic data is obtained using an advanced analytical model in MATLAB, based on the grid and radius method coupled with the Morgenstern Price method of slope analysis to identify FOS corresponding to the critical slip surface. Six ML models—multiple linear regression , support vector regression, XGBoost, random forest (RF), Gaussian process regression (GPR), and artificial neural network (ANN)—are evaluated and compared to identify the best model for FOS prediction. Among these, the GPR model exhibits superior performance, with <i>R</i><sup><i>2</i></sup> values of 99.8%, 99.7%, and 99.4% for training, validation, and testing datasets, respectively. The mean absolute error metrics are 1.0%, 1.2%, and 1.6%, and the mean squared error metrics are 0.3%, 0.5%, and 1.2%, respectively, demonstrating the model’s robustness on unseen geotechnical slope data. For slope failure classification, five ML models—logistic regression (LR), support vector machine (SVM), RF, XGBoost, and ANN—are utilized. The ANN model exhibited the highest performance metrics with a perfect recall (100%), as well as the highest F1-score (99.9%), specificity (98.1%), precision (99.8%), and accuracy (99.8%), making it the most effective model for this task. The results indicate that while all models perform well, the ANN and SVM models exhibit the highest overall performance metrics. The ANN model’s superior performance in the rank analysis demonstrates its robustness and reliability for slope failure detection. A modified representation of the confusion matrix is also presented in the current work for a reader-friendly visualization of the results of each model based on the precision, recall, F1-score, specificity, and accuracy metrics. The study also emphasizes the importance of interpretability by using Shapley Additive Explanations to ensure transparency and practical applicability. This approach advances data-centric geotechnics and provides reliable predictive capabilities suited to real-world geological conditions with fluctuating moisture and complex soil–water interactions.</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

Detection and prediction of slope stability in unsaturated finite slopes using interpretable machine learning

  • 期刊论文

  • 发表日期：

    2025年6月4日

查看全文
  • Kenue Abdul Waris,
  • Mohammed Asif ur Rahaman,
  • B. Munwar Basha

摘要

This study aims to enhance the domain of slope stability assessments by integrating unsaturated soil mechanics into machine learning (ML) methodologies. It addresses both regression and classification problems to predict the factor of safety (FOS) without the need to determine the critical slip surface iteratively. Orthogonal Latin hypercube sampling is employed to generate a comprehensive dataset of 16,219 data points. This synthetic data is obtained using an advanced analytical model in MATLAB, based on the grid and radius method coupled with the Morgenstern Price method of slope analysis to identify FOS corresponding to the critical slip surface. Six ML models—multiple linear regression , support vector regression, XGBoost, random forest (RF), Gaussian process regression (GPR), and artificial neural network (ANN)—are evaluated and compared to identify the best model for FOS prediction. Among these, the GPR model exhibits superior performance, with R2 values of 99.8%, 99.7%, and 99.4% for training, validation, and testing datasets, respectively. The mean absolute error metrics are 1.0%, 1.2%, and 1.6%, and the mean squared error metrics are 0.3%, 0.5%, and 1.2%, respectively, demonstrating the model’s robustness on unseen geotechnical slope data. For slope failure classification, five ML models—logistic regression (LR), support vector machine (SVM), RF, XGBoost, and ANN—are utilized. The ANN model exhibited the highest performance metrics with a perfect recall (100%), as well as the highest F1-score (99.9%), specificity (98.1%), precision (99.8%), and accuracy (99.8%), making it the most effective model for this task. The results indicate that while all models perform well, the ANN and SVM models exhibit the highest overall performance metrics. The ANN model’s superior performance in the rank analysis demonstrates its robustness and reliability for slope failure detection. A modified representation of the confusion matrix is also presented in the current work for a reader-friendly visualization of the results of each model based on the precision, recall, F1-score, specificity, and accuracy metrics. The study also emphasizes the importance of interpretability by using Shapley Additive Explanations to ensure transparency and practical applicability. This approach advances data-centric geotechnics and provides reliable predictive capabilities suited to real-world geological conditions with fluctuating moisture and complex soil–water interactions.