Estimation of Soil Aggregate Stability in Forest's Soils of Guilan Province by Artificial Neural Networks and Regression Pedotransfer Functions

Using artificial neural networks (ANNs) and regression pedotransfer functions to predict the surrogate soil properties such as aggregate stability reduces time and cost needed for their direct measurements. In this research، 100 soil samples were collected from the forest soils of Guilan province. Organic matter، bulk density، equivalent carbonate calcium، particle density، porosity، soil mechanical resistance، clay، sand، silt، pH and electrical conductivity all were measured as independent variables. Geometric mean diameter (GMD) was computed as dependent variable by appropriate method. The samples were divided into two data subsets randomly: 80 for model calibration and 20 for model test. Regression pedotransfer functions were generated by stepwise method. For establishing ANNs we used Marquardt-Levenburg training algorithm and a 3-layer perceptron structure with 6 neurons in one hidden layer. According to the correlation matrix between GMD as dependent variable and independent variables، 10 groups input variables were selected. The were employed once by multi-variate regression pedotransfer functions and once by artificial neural networks. According to the adjusted coefficient of determination (R2ady)، root mean square error (RMSE) and relative improvement (RI) a model resulted from applying ANNs and using input variables of pH، particle density، silt and soil mechanical resistance turned to be the best model for predicting GMD of the examined soils.