Estimation of Soil Infiltration in Agricultural and Pasture Lands using Artificial Neural Networks and Multiple Regressions

Common methods to determine the soil infiltration need extensive time and are expensive. However, the existence of non-linear behaviors in soil infiltration makes it difficult to be modeled. With regards to the difficulties of direct measurement of soil infiltration, the use of indirect methods toestimate this parameter has received attention in recent years. Despite the existence of various theoreticaland experimental equations, other indirect methods such as artificial neural networks are used to estimate this soil phenomenon. Currently, artificial neural networks are highly efficient in modeling non-linear equations. In the present studywe have collected 200 soil samples from Ghoshe location in Semnan Province. Half of samples were collected from agricultural lands and the other half were collected from nearby pasture lands. Some soil chemical as well as physical properties such as electrical conductivity (EC), soil texture, lime percentage, sodium adsorption ration (SAR) and bulk density were considered to be easy, rapidly obtainable features whereas soil permeability was considered to be a difficult, time consuming feature. The collected data were randomly divided in two categories of training and testing to be used for two artificial neural networks, multi-layer perception using back-propagation algorithm (MLP/BP).radial basis functions (RBF) and a nonlinear regression model. Results of this research showedthe high efficiency of the artificial neural network compared with multiple regression. The MLP network was better than the RBF network. Sensitive analysis was used to determine the correlation between independent parameters and permeability.