PEDOTRANSFER FUNCTION FOR DETERMINING SATURATED HYDRAULIC CONDUCTIVITY USING ARTIFICIAL NEURAL NETWORK (ANN)

In the work there were presented two pedotransfer models for determination of saturated hydraulic conductivity, generated by artificial neural networks (ANN). Models were learned based on empirical data obtained in laboratory, on 56 soil samples of differentiated texture. In the first model the input parameters were: characteristic diameters d(10), d(50), d(60), d(90), content of sand, silt and clay fractions, total porosity, bulk density and organic matter content. The MLP type of ANN was used. The best fitted model turned out MLP 10-10-1 with satisfactory quality parameters, for learning 0,996, for testing 0,754 and for validation 1,000. Global sensitivity analysis showed that the highest influence on explanation of relationship between saturated hydraulic conductivity in this model had: clay content (absolute influence 37.7%, d(60) (17.1%), sand content (13.5%), d(90) (6.0%), bulk density (5.9%) and total porosity (5.7%). The remaining parameters had absolute influence below 5.0%). The next generated ANN model was MLP 6-10-1, with six explaining parameters, of greatest influence. Correlation coefficient attained value 0.989 and 0.955 for the first and the second model. Mean percentage error pointed out underestimation in comparison to laboratory measurement. The values attained 35.9% and 54.8% respectively. Limitation of explaining parameters did not point high deterioration of the ANN model quality.