Predicting the cone penetration resistance of unsaturated agricultural soils by neural network

This study develops and evaluates an Artificial Neural Network (ANN) model for predicting cone penetration resistance (CPR) in unsaturated silt loam soils. Controlled soil bin experiments were conducted under varying moisture contents and compaction levels to investigate the relationships among dry density, moisture content, and CPR, and a three-dimensional representation of these interactions was generated. The experimental data were used to train an ANN with the Levenberg–Marquardt algorithm, and model robustness was ensured through fivefold cross-validation. The predictive performance of the ANN was benchmarked against Support Vector Machine (SVM), Gaussian Process Regression (GPR), and Random Forest (RF) models. Results demonstrated that the ANN outperformed the alternative approaches, achieving a cross-validated coefficient of determination (R²) of 0.95, a root mean square error (RMSE) of 125.82 kPa, and mean absolute error (MAE) of 74.49 kPa. Feature importance analysis identified dry density as the most influential predictor, consistent with established soil mechanics principles. The findings highlight the potential of ANN models as reliable tools for predicting soil penetration resistance, with practical implications for precision agriculture and soil compaction management. While this study focused on a single soil type, future research will extend the approach to a broader range of soil textures.