Estimating soil water infiltration parameters using Artificial Neural Networks

Infiltration is a significant process which controls the fate of water in the hydrologic cycle. The direct measurement of infiltration is time consuming, expensive and often impractical because of the large spatial and temporal variability. Artificial Neural Networks (ANNs) are used as an indirect method to predict the hydrological processes. The objective of this study was to develop and verify some ANNs to predict the infiltration process. For this purpose, 123 double ring infiltration data were collected from different sites of Iran. The parameters of some infiltration models were then obtained; using sum squares error optimization method. Basic soil properties of the two upper pedogenic layers such as initial water content, bulk density, particle-size distributions, organic carbon, gravel content, CaCO3 percent and soil water contents at field capacity and permanent wilting point were obtained for each sampling point. The feedforward multilayer perceptron was used for predicting the infiltration parameters. Two ANNs types were developed to estimate infiltration parameters. The developed ANNs were categorized into two groups; type 1 and type 2 ANNs. For developing type 1 ANNs, the basic soil properties of the first upper soil horizon were used as inputs, hierarchically. While for developing type 2 ANNs the basic soil properties of the two upper soil horizons were used as inputs, using principal component analysis technique. Evaluation results of these two types ANNs showed the better performance of type 1 ANNs in predicting the infiltration parameters. Therefore, this type of ANNs was used for predicting the cumulative infiltration. The reliability test indicated that the developed ANNs for Philip model have the best performance to predict cumulative infiltration with a mean RMSE of 6.644 cm. The developed ANNs for Horton, Kostiakov-Lewis and Kostiakov have the next best ranks, respectively.