Abstract: Although it is essential to determine saturated hydraulic conductivity to investigate solute movement in soil, drainage systems design, determination of rate of water table falling and infiltration rate, but its measurement is time-consuming, difficult and expensive. Thus many efforts have been made to indirectly estimate hydraulic conductivity and other costly measured properties via establishing relationship between this properties and readily available data. In this research, first clay%, sand%, silt%, organic carbon% and bulk density were determined as readily available data and then saturated hydraulic conductivity was also using double-ring method in 221 samples of clay soil in Guilan province. Artificial neural network (ANN) with 3 layer proceptron structure and Marquardt-Levenberg training algorithm is established by NeuralWorks Pro II Plus software. Number of hidden layer and its neurons were determined by try and error method and equal to 1 and 6 respectively.187 data subset were used for training and calibration of neural networks and other 34 data subset used for neural network test. Three patterns of input readily available data were utilized: SSC (silt, sand and clay content), CBDOC (Clay, bulk density and organic carbon) and SSCBDOC (all of parameters). ANN with CBDOC input pattern with R2=0.974 and RMSR = 0.160 showed the most accurate prediction for saturated hydraulic conductivity. The CBDOC relative improvement (RI) to SSCBDOC and SSC was equal to %2.4 and %48. Results showed that artificial neural network with 3-layer proceptron structure, one hidden layer and six neuron in it and CBDOC input pattern could be used to predict saturated hydraulic conductivity.

Keywords: Artificial neural networks, Clay soils, Modeling, Saturated hydraulic conductivity.