Abstract
Cation exchange capacity (CEC) of soil is an important characteristic that determines the soil buffering capacity to maintain cationic nutrients and pollutants against leaching to the subsurface layers. Radial basis function neural networks (RBFN) were applied into the estimation of cation exchange capacity (CEC) from soil physicochemical properties. A total of 457 soil samples collected from surface to deep underground horizons under various climatic and geographic conditions were successfully compiled in the comprehensive modelling research. Soil horizon; pH; organic carbon content; and clay, silt, and sand contents were taken as the input variables. Four RBFN models applied in CEC estimates for A, B, C horizons as well as the aggregate soils from all horizons were developed, with high correlation coefficients of 0.9397, 0.8368, 0.8014, and 0.8335 between the predicted and measured values of the external test samples, respectively, after optimization of the network architectures. Furthermore, the influence of each input variable on the soil CEC was particularly evaluated and sequenced by sensitivity analysis, confirming that clay content, organic carbon content, and soil horizons are the most important determinants of CEC estimate. Due to the existing nonlinear and uncertain characteristics of natural soil properties demonstrated in this paper, the optimized RBFN models showed high superiority in large-scale data simulation of complicated soil samples compared with multiple linear regressions (MLR). All the results confirm that an artificial neural network is a precise, sensitive, and robust method in the quantitative study of CEC in widely distributed soil samples in the complex ecosystem.