Evaluation of standard, black-box, and bayesian RSM-SVR models in the semi-arid area of south-eastern Iran for predicting soil chemical properties

This study evaluates three hybrid modeling approaches combining Response Surface Methodology and Support Vector Regression (RSM-SVR) for forecasting soil chemical properties. The study focuses on semiarid regions of Iran with coarse-textured soils. Exchangeable Sodium Percentage (ESP), Cation Exchange Capacity (CEC), and Sodium Adsorption Ratio (SAR) were predicted using three modeling approaches: Standard, Black‑Box, and Bayesian RSM‑SVR. The models were developed from 258 soil samples. These samples were thoroughly characterized and identified as having sandy loam and sand textures. Black-Box RSM-SVR was more suitable for ESP prediction (R2 = 0.99, RMSE = 0.92) and CEC prediction (R2 = 0.95, RMSE = 1.58) compared to the Bayesian approach, while the Bayesian approach performed a bit better in SAR prediction (R2 = 0.90, RMSE = 2.62). Feature importance analysis indicated sand content as the primary driver for CEC prediction (0.68 cmol/kg), while silt controlled sodium-related parameter predictions (ESP: 0.32%, SAR: 0.38). All models produced lower accuracy when forecasting extreme values corresponding to extremely degraded soils. In comparison with the Standard method, Black-Box methodology reduced prediction errors by approximately 40% for ESP and 28% for CEC in soils with prevailing sandy textures. These findings add to the understanding of forecasting soil chemical characteristics in challenging semiarid situations and offer applicable knowledge for precision agriculture and sustainable land use management in sodic and resource-scarce areas.