Modelling soil cation exchange capacity in different land-use systems using artificial neural networks and multiple regression analysis

Cation exchange capacity (CEC), as an important indicator of soil quality, represents the ability of the soil to hold positively charged ions. In this study, CEC was successfully predicted using different statistical methods, including artificial neural networks (ANNs) involving multi-layer perceptron (MLP), radial basis function (RBF), multiple linear regression (MLR) and nonlinear regression (NLR). About 293 soil samples were collected from North East India, which are under three land uses (shifting agriculture (jhum), forest and cash crops). Also, 70% of the samples (205 samples) was selected as the calibration set and the remaining 30% (88 samples) used as the prediction set. Soil pH, texture, bulk density (BD) and organic carbon (OC) were used as predictor variables to estimate CEC. The CEC-pedotransfer function (CEC-PTF) performance was evaluated with the coefficient of determination (R-2), root mean square error (RMSE) and standard error for the estimate (SEE) between the observed and predicted values. The results indicated that the nonlinear model (R-2 = 0.91 and SEE = 1.82 for training) for cash-crop system, and RBF (R-2 = 0.91 and SEE = 3.83 for training) for jhum system were the best models to estimate CEC. In contrast, RBF (R-2 = 0.67 and SEE = 14.87 for training) for forest system was the worst model to estimate CEC. The results confirm that clay and OC were the most influential variables to predict CEC in the cash-crop system, whereas BD and OC were more suitable for jhum system. Although the ANNs provided suitable predictions of the entire dataset, NLR gave a formula to estimate soil CEC using commonly tested soil properties. Thus, NLR provided a reasonable estimate of CEC for most soils analysed.