Estimating of Field Capacity, Permanent Wilting and Available Water Content in Ardabil Plain Soils using Regression and Artificial Neural Network Models

Direct measurement of soil hydraulic properties is consuming, costly and sometimes unreliable because of soil heterogeneity and experimental errors. These properties can be estimated from surrogate data such as particle size distribution, bulk density, organic carbon and CaCO3 using pedotransfer functions (PTFs). The objective of this research was presentation of regression and neural network models for estimation of missing soil properties including field capacity, permanent wilting and available water contents from above-cited surrogate soil properties in some soil of the Ardabil Plain. Total 100 soil samples were taken and then some physical and chemical properties of them measured. Soil samples were divided into two groups as 80 for the development and 20 for the validation of PTFs. Neural network and regression models were made using Neurosolution5 and SPSS softwares, respectively. The values of determination coefficient (R2) and root mean square error (RMSE) for the estimation of field capacity, permanent wilting and available water contents were obtained 0.82 and 2.29, 0.82 and 1.38, 0.57 and 1.97 in the best regression models and 0.87 and 1.9, 0.90 and 1.02, 0.73 and 1.56 in the best neural network models, respectively. The values of R2 and RMSE for the results of regression and artificial neural network PTFs showed that both models can be applied to predicting missing soil properties. Regression models hadn’t any efficiency to predict available water content. Artificial neural networks were performed better than regression models in this case.