Machine Learning-Based Prediction of Soil Electrical Resistivity Using Field-Instrumented Hydrologic Data

Subsurface characterization is challenging due to soil heterogeneity and the complex interplay of moisture, suction, and temperature. The ability to deduce these variables through a single measurable soil property, specifically soil resistivity, can potentially enhance subsurface characterization. This study aimed to investigate the relationship between field-surveyed soil electrical resistivity and basic soil physical properties to develop a predictive model using various machine learning-based regression techniques. To achieve this objective, moisture and temperature sensors, along with tensiometers, were installed at different depths in several vegetated soils, and resistivity tests were conducted periodically at the same locations. Two regularized linear regression models, Ridge (L2 penalty) and Lasso (L1 penalty), were then developed and compared to predict bulk resistivity from the three hydrological variables. All predictors were standardized before analysis, and hyperparameters were optimized via five-fold cross-validation. The performance of the two regression techniques was evaluated using the metrics Root Mean Square Error (RMSE) and Coefficient of Determination (R²). Results obtained from the study indicated that Ridge regression achieved the highest predictive accuracy (R² = 0.778; RMSE = 14.28 Ω-m). Conversely, Lasso regression yielded a sparser model (R² = 0.692; RMSE = 19.40 Ω-m). The results also showed that soil suction has negligible contributions in correlating soil resistivity with soil physical properties, especially in Ridge regression. These results demonstrate that, among the tested methods, Ridge regression most reliably captures the linear relationships between hydrological state variables and soil resistivity in vegetated soil.