Scholarlink

<p>Soil electrical resistivity testing is a rapid and non-destructive method for geotechnical site characterization, but conventional regression approaches often struggle to capture its nonlinear behavior. This study evaluates three machine learning models—Support Vector Regression (SVR), Random Forest (RF), and Extreme Gradient Boosting (XGB)—for predicting soil resistivity using water content, plasticity index, and dry density. A total of 30 soil samples were collected from power grid substations in three regions of Thailand and tested under controlled laboratory conditions. Each sample was measured at five time intervals (Day 0, 1, 7, 14, and 28), resulting in a dataset of 150 observations that captures temporal variations in moisture and resistivity. Model performance was assessed using R² and RMSE across training, testing, and 5-fold cross-validation, with resistivity values log-transformed prior to model training. SVR showed the most consistent performance, with R² values of 0.938, 0.853, and 0.813 for training, testing, and cross-validation, respectively. XGB also performed well but showed reduced generalization on unseen data, while RF consistently yielded lower accuracy. Permutation importance analysis identified water content as the dominant predictor, followed by plasticity index and dry density. Prediction errors increased at higher resistivity values, particularly under dry conditions, reflecting increased nonlinearity and data variability. Overall, SVR provides a reliable approach for predicting soil resistivity from basic geotechnical parameters, especially for small to medium-sized, time-dependent datasets.</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

Predictive modeling of soil electrical resistivity using ensemble machine learning algorithms with geotechnical parameters

  • 期刊论文

  • 发表日期：

    2026年5月15日

查看全文
  • Kornkanok Sangprasat,
  • Avirut Puttiwongrak,
  • Shinya Inazumi

摘要

Soil electrical resistivity testing is a rapid and non-destructive method for geotechnical site characterization, but conventional regression approaches often struggle to capture its nonlinear behavior. This study evaluates three machine learning models—Support Vector Regression (SVR), Random Forest (RF), and Extreme Gradient Boosting (XGB)—for predicting soil resistivity using water content, plasticity index, and dry density. A total of 30 soil samples were collected from power grid substations in three regions of Thailand and tested under controlled laboratory conditions. Each sample was measured at five time intervals (Day 0, 1, 7, 14, and 28), resulting in a dataset of 150 observations that captures temporal variations in moisture and resistivity. Model performance was assessed using R² and RMSE across training, testing, and 5-fold cross-validation, with resistivity values log-transformed prior to model training. SVR showed the most consistent performance, with R² values of 0.938, 0.853, and 0.813 for training, testing, and cross-validation, respectively. XGB also performed well but showed reduced generalization on unseen data, while RF consistently yielded lower accuracy. Permutation importance analysis identified water content as the dominant predictor, followed by plasticity index and dry density. Prediction errors increased at higher resistivity values, particularly under dry conditions, reflecting increased nonlinearity and data variability. Overall, SVR provides a reliable approach for predicting soil resistivity from basic geotechnical parameters, especially for small to medium-sized, time-dependent datasets.