Study of Pedotransfer Functions Multivariate regression, MLP and RBF to estimate CEC for Soils of North Ahvaz

To estimate the Cation Exchange Capacity (CEC), indirect manner used of Pedotransfer Functions (PTFs). CEC is one of the important soil fertility factors, and not measured directly because it is costly and time consuming. Thus, used from regression equations between easily and non-easily soil properties. The purpose of this research, is develop the PTFs for CEC, with use of easily available soil properties. For this purpose, measured for 100 sample of soil contain of 1000 data include soil particle size distribution, bulk density, organic matter, lime, pore space, geometric mean diameter and geometric standard deviation were done. After data normalization, were done PTFs with Multivariate Regression (MR) in SPSS and Artificial Neural Networks (ANNs) for soil CEC in MATLAB software. The ANNs used in research are Multi-Layer Perceptron (MLP) and Radial Basis Function (RBF). Education of ANNs are based on trial and error, until arrived suitable outputs with changes hidden layer number and neuron number. From all data were Selected 70% (700) data for training and 30% (300) for teste. Then were entered all data to software and test different networks with one hidden layer. The network was design with trial and error to maximum correlation coefficient and minimum mean square error (MSE). The results were shows MR is suitable for predict CEC (R2 =0.87) and for MLP network, were shows ANN can good estimated CEC with used of easily soil properties. MLP network able to estimate CEC with 9 neurons in input layer, 7 neurons in hidden layer and 1 neuron in output layer with tangent sigmoid transfer function, Linear transfer function and Bayesian learning algorithm with coefficient correlation 0.97 and MSE 0.013. For RBF networks to estimation CEC coefficient correlation and MSE were 0.55, 0.017 respectively. Results shows the MLP network with 1 hidden layer for estimation CEC with use of soil distribution, bulk density, organic matter and lime, is better than RBF network compared with of MSE and coefficient correlation.