Effect of type and count of input variables on moisture retention curve and saturated hydraulic conductivity prediction

Pedotransfer functions (PTFS) are useful means of prediction many properties of the soil, and especially the hydraulic characteristics of this porous media. The main advantages of this functions, as compare to conventional methods used to directly estimate soil hydraulic properties, is that they are not time-cost consuming. Different approaches such as classic linear and non linear regressions, artificial neural networks and regressions tree are being employed to develop the PTFS. Rosetta is a software package to predict soil hydraulic properties making use of artificial neural networks- based PTFS. In the present study, the impacts of the type and count of input variables to this software, on the prediction of the moisture retention curve and saturated hydraulic conductivity were evaluated in some soils from northern region of Iran, classed as of Loam and Clay Loam textures (USDA). Our results indicated that addition of bulk density as input variable decreased the performance of moisture retention curve prediction in both textural classes. Addition of bulk density showed on RMSE, ME, GMER and GSDER a positive and negative effect in Loam and Clay Loam textures, respectively. Addition of one or two moisture retention point(s) (the moisture content at matric potential of -33 and -1500 kpa) significantly decreased the RMSE at the medium range of matric potential (i.e. -33 to -500 kpa) and especially at -33 kpa. All of the studied PTFS tended to underestimate both saturated hydraulic conductivity and moisture content at different matric potential.