A comparison of metaheuristic optimizations with automated hyperparameter tuning methods in support vector machines algorithm for predicting soil water characteristic curve

Estimating the soil water characteristic curve (SWCC) is essential for understanding soil behavior related in geological and geotechnical, and environmental engineering. This study was designed to evaluate and compare metaheuristic methods (cuckoo search optimization (CSO) and grey wolf optimization (GWO)) with automated methods (Bayesian optimization (BO) and grid search (GS)) for tuning hyperparameters (penalty coefficient (C), insensitive loss (ε), and kernel width (γ)) in support vector machines (SVM) to improve SWCC estimation. Four pedotransfer functions (PTFs) were derived to estimate the parameters of the Brutsaert model using various input variables such as sand, clay, and bulk density (BD), as well as moisture content at 33 (FC) and 1500 kPa (PWP) from 354 UNSODA soil samples. The findings of the testing phase indicated that the BO-based SVM algorithm outperformed other optimization methods with an average error value of 0.057 cm³cm⁻³ for all PTFs. In PTF4 (sand+clay+BD + FC + PWP), BO demonstrated 6.23 %, 10.53 %, and 12.96%96 % higher reliability than CSO, GS, and GWO, respectively. The Shapley additive explanations analysis indicated that C parameter had the highest impact on model reliability, while ε parameter had the lowest. Finally, the integration of BO into SVM can improve accuracy, efficiency, and robustness in SWCC estimation, providing more reliable predictions for future geotechnical and hydrological studies.