Application of Radial Basis Function Neural Networks for prediction of sorptivity coefficient in calcareous soils

Sorptivity coefficient (S) is one of the most important parameter in infiltration and hydrological models. On one hand, its laboratory or field determination is usually labor-intensive, timeconsuming and expensive. On the other hand, due to their high inherent spatio-temporal variability, large number of samples is needed for its determination. Therefore, most investigators prefer to predict the hydraulic parameters by pedotransfer functions, PTFs, artificial neural networks ANNs using the easily available soil data. Recently, Radial Basis Function Neural Networks (RBFNNs) have been widely used for classification and function approximation. To the best of our knowledge, this type of ANN have been not used for predition of S parameter especiaaly for calcareous soils. Therefore, the objective of this study was to predict the S parameter of calcareous soils using RBFNNs. In this study 47 data (75% of the total dataset) for the six measured soil attributes of EC, pH, Wi, BD, MWD and GMD were considered as input variables in training the applied RBFNNs to predict the S parameter. Furthermore, 15 data (25% of the total dataset) of the aforementioned soil attributes were used to test the designed networks. For both training and testing steps the measured S values were compared with their corresponding RBFNNs-predicted values. Results indicate the R values in prediction of S parameter for the total, training and testing datasets are 0.979, 0.996 and 0.824, respectively. The amount of RMSE for the mentioned datasets was 0.0176, 0.0082, 0.0328 and MAPE for the mentioned datasets was 1.9825, 0.9221 and 5.3049, respectively. The training and testing datasets showed the minimum and the maximum deviation from the line of 1:1 and the mentioned deviation was moderate for that of the total dataset. Generally, since the excellent reliable predictions were obtained for the S parameter, it is recommended that such intelligence models instead of costly and time consuming procedures can be applied to predict S that is one of the vita hydraulic parameters in almost all of the infiltration and hydrological modeling. As it was mentioned earlier, for applying such an easy prediction, only the textural components, bulk density, initial water content, electrical conductivity and pH, those are the easily available attributes from the general soil analysis, are required. As Moosavi and Sepaskhah (2012a) concluded the tentative conclusion is that the proposed network may perform differently for other soil conditions, input variable, or for these proposed input variables outside the range of parameter values presented here, which may result in less accurate estimations.

Keywords:

Artificial neural networks , sorptivity coefficient , calcareous soils , correlation coefficient