A Comparison of Different Methods of Developing Pedotransfer Functions in Soils of Humid Regions in Iran

The functions employed in an estimation of costly measured soil properties from either widely available or more easily obtained basic soil properties are referred to as pedotransfer functions. To develop pedotransfer functions, one can use multivariate regression, neural networks and neuro-fuzzy models. To make a comparison among the mentioned models, 153 soil samples were collected from soils in Rasht Province. Clay, sand, silt as well as organic carbon percentage considered as readily obtainable parameters vs. cation exchange capacity as predicted variable were assessed. The data set was broken into two subsets for calibration (80%) and testing (20%) of the models. According to some such evaluation parameters as Root Mean Square, Average Error and Coefficient of Determination, neuro-fuzzy benefited from the most accuracy for a prediction of cation exchange capacity. Also, results indicated that the neuro-fuzzy model increased the accuracy of cation exchange capacity prediction for about 14%. Following neuro-fuzzy model, artificial neural network (Feed forward, General Regression Neural Network, and Cascade Forward) benefited from higher accuracies than Multivariate Regression approach.