A Sensitivity Analysis of ANN Pedotransfer Functions for spatial modeling of Soil Cation Exchange Capacity

The development of models simulating soil processes has increased rapidly in recent years. These models havebeen developed to improve the understanding of important soil processes and also to act as tools for evaluatingagricultural and environmental problems. In this research, an artificial neural network (ANN) model was developed topredict of soil Cation Exchange Capacity (CEC) which was called neural kriging (NK) by easily measurablecharacteristics of clay and organic carbon. 134 soil samples were collected from different horizons of 34 soil profileslocated in the Ziaran region, Qazvin province, Iran. The data set was divided into two subsets for calibration (75%) andtesting (25%) of the model. In order to evaluate the model, root mean square error (RMSE) and R2 were used. The valueof RMSE and R2 derived by ANN model were 0.04 and 0.97, respectively. The comparison of RMSE and R2 forvarious ANN models showed that the ANN model with three neurons in hidden layer gives better estimates of soilCEC. Sensitivity analysis was also conducted to investigate the effects of various explanatory parameters on the output.The results indicated that CEC variation was more sensitive to clay content than OC variable. For geostatisticalanalyzing, sampling was done with stratified random method and 34 soil samples from 0 to 15 cm depth were collectedwith auger within 34 locations. For comparing and evaluation of neural kriging and ordinary kriging methods, crossvalidation was used by statistical parameters of RMSE and correlation coefficient (r) for test data set. The resultsshowed that neural kriging method has the higher correlation coefficient (0.96) and less RMSE (1.22) than ordinarykriging method in predicting and spatial mapping of soil CEC in unsampled areas.